diff options
| author | JRPan <[email protected]> | 2023-06-17 10:44:14 +0800 |
|---|---|---|
| committer | GitHub <[email protected]> | 2023-06-17 10:44:14 +0800 |
| commit | 46428eeaf3cfdb87fb6c323afd92e4bd79c54476 (patch) | |
| tree | 144c661d4b11c28bb02caf2dcbc45cf8e70d7838 | |
| parent | 4f1cdbe1ff00ab003cc9d33f1a6943eeb6b203a3 (diff) | |
| parent | 66ae8815dd1a9cd5fcc0e08332ebca323c356863 (diff) | |
Merge branch 'dev' into dev
| -rw-r--r-- | configs/tested-cfgs/SM7_GV100/accelwattch_ptx_sim.xml | 623 | ||||
| -rw-r--r-- | configs/tested-cfgs/SM7_GV100/accelwattch_ptx_sim_alt.xml | 623 | ||||
| -rw-r--r-- | configs/tested-cfgs/SM7_GV100/accelwattch_sass_hw.xml | 613 | ||||
| -rw-r--r-- | configs/tested-cfgs/SM7_GV100/accelwattch_sass_hybrid.xml | 613 | ||||
| -rw-r--r-- | configs/tested-cfgs/SM7_GV100/accelwattch_sass_sim.xml | 613 | ||||
| -rw-r--r-- | configs/tested-cfgs/SM7_GV100/accelwattch_sass_sim_alt.xml | 613 | ||||
| -rw-r--r-- | configs/tested-cfgs/SM7_GV100/config_volta_islip.icnt | 74 | ||||
| -rw-r--r-- | configs/tested-cfgs/SM7_GV100/gpgpusim.config | 237 | ||||
| -rw-r--r-- | configs/tested-cfgs/SM7_QV100/gpgpusim.config | 2 |
9 files changed, 4010 insertions, 1 deletions
diff --git a/configs/tested-cfgs/SM7_GV100/accelwattch_ptx_sim.xml b/configs/tested-cfgs/SM7_GV100/accelwattch_ptx_sim.xml new file mode 100644 index 0000000..d94d5bd --- /dev/null +++ b/configs/tested-cfgs/SM7_GV100/accelwattch_ptx_sim.xml @@ -0,0 +1,623 @@ +<?xml version="1.0" ?>
+<component id="root" name="root">
+ <component id="system" name="system">
+<!-- Copyright (c) 2018-2021, Vijay Kandiah, Junrui Pan, Mahmoud Khairy, Scott Peverelle, Timothy Rogers, Tor M. Aamodt, Nikos Hardavellas
+Northwestern University, Purdue University, The University of British Columbia
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+ list of conditions and the following disclaimer;
+2. Redistributions in binary form must reproduce the above copyright notice,
+ this list of conditions and the following disclaimer in the documentation
+ and/or other materials provided with the distribution;
+3. Neither the names of Northwestern University, Purdue University,
+ The University of British Columbia nor the names of their contributors
+ may be used to endorse or promote products derived from this software
+ without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE. -->
+
+ <!-- DYNAMIC Power Activity Factors -->
+
+ <param name="TOT_INST" value="2" /> <!--Instruction Buffer Activity Factor(Total Warp Instructions)-->
+ <param name="FP_INT" value="4.57" /> <!--Scheduler Activity Factor(Non-Memory Warp Instructons)-->
+ <param name="IC_H" value="11.44089762" /> <!--Instruction Cache Hits -->
+ <param name="IC_M" value="21.76302498" /> <!--Instruction Cache Misses-->
+ <param name="DC_RH" value="7.737353491" /> <!--L1 Data Cache Read Hits-->
+ <param name="DC_RM" value="8.618027871" /> <!--L1 Data Cache Read Misses-->
+ <param name="DC_WH" value="0.53469516" /> <!--L1 Data Cache Write Hits-->
+ <param name="DC_WM" value="13.9055689" /> <!--L1 Data Cache Write Misses-->
+ <param name="CC_H" value="0.65916315" /> <!--Constant Cache Hits-->
+ <param name="CC_M" value="0.73418985" /> <!--Constant Cache Misses-->
+ <param name="SHRD_ACC" value="1.313660815" /> <!--Shared Memory Accesses-->
+ <param name="REG_RD" value="0.053279375" /> <!--Register Reads-->
+ <param name="REG_WR" value="0.079919063" /> <!--Register Writes-->
+ <param name="INT_ACC" value="3.429666768" /> <!--Integer ALU Accesses-->
+ <param name="FP_ACC" value="0.711591276" /> <!--FPU Accesses-->
+ <param name="DP_ACC" value="0.742812382" /> <!--Integer Accesses-->
+ <param name="INT_MUL_ACC" value="0.148636575" /> <!--Integer MUL Accesses-->
+ <param name="FP_MUL_ACC" value="0.212559571" /> <!--FP MUL Accesses-->
+ <param name="FP_SQRT_ACC" value="1.241271438" /> <!--SFU: sqrt Accesses-->
+ <param name="FP_LG_ACC" value="0.59034036" /> <!--SFU: log Accesses-->
+ <param name="FP_SIN_ACC" value="0.212555149" /> <!--SFU: sin/cos Accesses-->
+ <param name="FP_EXP_ACC" value="0.702043615" /> <!--SFU: exponent Accesses-->
+ <param name="DP_MUL_ACC" value="0.282564496" /> <!--DP MUL Accesses-->
+ <param name="TENSOR_ACC" value="2.485" /> <!--Tensor Accesses-->
+ <param name="TEX_ACC" value="0.212559047" /> <!--Texture Unit Accesses-->
+ <param name="MEM_RD" value="0.02772" /> <!--DRAM Reads-->
+ <param name="MEM_WR" value="0.0336" /> <!--DRAM Writes-->
+ <param name="MEM_PRE" value="0.00924" /> <!--DRAM Precharge-->
+ <param name="L2_RH" value="1.046834662" /> <!--L2 Data Cache Read Hits-->
+ <param name="L2_RM" value="2.670605032" /> <!--L2 Data Cache Read Misses-->
+ <param name="L2_WH" value="3.269555394" /> <!--L2 Data Cache Write Hits-->
+ <param name="L2_WM" value="2.18020968" /> <!--L2 Data Cache Write Misses-->
+ <param name="NOC_A" value="83.18977901" /> <!--Interconnect Network Accesses-->
+ <param name="PIPE_A" value="0.0257" /> <!--Pipeline Accesses-->
+
+ <!-- PTX Model specific DYNAMIC Power Activity Factors-->
+
+ <param name="TC_H" value="0.01021" /> <!--Texture Cache Hits-->
+ <param name="TC_M" value="0.02466" /> <!--Texture Cache Misses-->
+ <param name="INT_MUL24_ACC" value="0.290770573" /> <!--Integer 24-bit MUL Accesses-->
+ <param name="INT_MUL32_ACC" value="0.252598514" /> <!--Integer 32-bit MUL Accesses-->
+ <param name="INT_DIV_ACC" value="5.121706665" /> <!--Integer DIV Accesses-->
+ <param name="FP_DIV_ACC" value="4.599926258" /> <!--FP DIV Accesses-->
+ <param name="DP_DIV_ACC" value="25.84876211" /> <!--DP DIV Accesses-->
+
+ <!-- STATIC & CONSTANT Power -->
+
+ <param name="constant_power" value="32.32522272" /> <!--Constant power-->
+ <param name="idle_core_power" value="0.28279166"/> <!--Idle SM power-->
+
+ <param name="static_cat1_flane" value="15.29035866"/> <!--INT (ADD+MUL) First Lane Activation Power-->
+ <param name="static_cat1_addlane" value="0.586233603"/> <!--INT (ADD+MUL) Additional Lane Activation Power-->
+
+ <param name="static_cat2_flane" value="18.6179906"/> <!--INT+FP First Lane Activation Power-->
+ <param name="static_cat2_addlane" value="0.645228013"/> <!--INT+FP Additional Lane Activation Power-->
+
+ <param name="static_cat3_flane" value="19.10017723"/> <!--INT+FP+DP First Lane Activation Power-->
+ <param name="static_cat3_addlane" value="0.726863055"/> <!--INT+FP+DP Additional Lane Activation Power-->
+
+ <param name="static_cat4_flane" value="18.55029744"/> <!--INT+FP+SFU First Lane Activation Power-->
+ <param name="static_cat4_addlane" value="0.6099397"/> <!--INT+FP+SFU Additional Lane Activation Power-->
+
+ <param name="static_cat5_flane" value="14.74826681"/> <!--INT+FP+TEX First Lane Activation Power-->
+ <param name="static_cat5_addlane" value="0.514367937"/> <!--INT+FP+TEX Additional Lane Activation Power-->
+
+ <param name="static_cat6_flane" value="48.94875596"/> <!--INT+FP+TENSOR First Lane Activation Power-->
+ <param name="static_cat6_addlane" value="0.0"/> <!--INT+FP+TENSOR Additional Lane Activation Power-->
+
+ <param name="static_light_flane" value="1.965373811"/> <!--LIGHT_SM First Lane Activation Power-->
+ <param name="static_light_addlane" value="0.003966868"/> <!--LIGHT_SM Additional Lane Activation Power-->
+
+ <param name="static_intadd_flane" value="19.70468506"/> <!--INT ADD First Lane Activation Power-->
+ <param name="static_intadd_addlane" value="0.388578623"/> <!--INT ADD Additional Lane Activation Power-->
+
+ <param name="static_intmul_flane" value="16.64811823"/> <!--INT MUL First Lane Activation Power-->
+ <param name="static_intmul_addlane" value="0.281803166"/> <!--INT MUL Additional Lane Activation Power-->
+
+ <param name="static_geomean_flane" value="17.21745077"/> <!--GEOMEAN First Lane Activation Power-->
+ <param name="static_geomean_addlane" value="0.650630555"/> <!--GEOMEAN Additional Lane Activation Power-->
+ <param name="static_shared_flane" value="31.40965691"/> <!--Shared Memory First Lane Activation Power-->
+ <param name="static_l1_flane" value="34.79491352"/> <!--L1D First Lane Activation Power-->
+ <param name="static_l2_flane" value="17.30654755"/> <!--L2D ADD+MUL First Lane Activation Power-->
+
+
+ <!-- LEAVE BELOW UNCHANGED! These are parameters used for initial GPUWattch model left untouched -->
+
+
+ <param name="GPU_Architecture" value="1"/><!-- 0-G80; 1-Fermi; others not supported -->
+ <param name="architecture" value="1"/> <!-- fermi:1 quadro:2 other: undefined-->
+ <param name="number_of_L1Directories" value="0"/>
+ <param name="number_of_L2Directories" value="0"/>
+ <param name="number_of_L2s" value="1"/> <!-- This number means how many L2 clusters in each cluster there can be multiple banks/ports -->
+ <param name="number_of_L3s" value="0"/> <!-- This number means how many L3 clusters -->
+ <param name="number_of_NoCs" value="1"/>
+ <param name="homogeneous_cores" value="1"/><!--1 means homo -->
+ <param name="homogeneous_L2s" value="1"/>
+ <param name="homogeneous_L1Directorys" value="1"/>
+ <param name="homogeneous_L2Directorys" value="1"/>
+ <param name="homogeneous_L3s" value="1"/>
+ <param name="homogeneous_ccs" value="1"/><!--cache coherece hardware -->
+ <param name="homogeneous_NoCs" value="1"/>
+ <param name="core_tech_node" value="23"/><!-- nm -->
+
+ <param name="number_cache_levels" value="2"/>
+ <param name="interconnect_projection_type" value="0"/><!--0: agressive wire technology; 1: conservative wire technology -->
+ <param name="device_type" value="0"/><!--0: HP(High Performance Type); 1: LSTP(Low standby power) 2: LOP (Low Operating Power) -->
+ <param name="longer_channel_device" value="1"/><!-- 0 no use; 1 use when possible -->
+ <param name="machine_bits" value="32"/>
+ <param name="virtual_address_width" value="32"/>
+ <param name="physical_address_width" value="32"/>
+ <param name="virtual_memory_page_size" value="4096"/>
+
+ <param name="IDLE_CORE_N" value="1"/>
+ <stat name="num_idle_cores" value="0"/><!-- Average Number of idle cores during this period -->
+ <stat name="total_cycles" value="total_cycles_match_mcpat"/>
+ <stat name="idle_cycles" value="idle_cycles_match_mcpat"/>
+ <stat name="busy_cycles" value="busy_cycles_match_mcpat"/>
+ <!--This page size(B) is complete different from the page size in Main memo secction. this page size is the size of
+ virtual memory from OS/Archi perspective; the page size in Main memo secction is the actuall physical line in a DRAM bank -->
+ <!-- *********************** cores ******************* -->
+ <component id="system.core0" name="core0">
+ <!-- Core property -->
+ <param name="clock_rate" value="700"/>
+ <param name="instruction_length" value="32"/>
+ <param name="opcode_width" value="9"/>
+ <!-- address width determins the tag_width in Cache, LSQ and buffers in cache controller
+ default value is machine_bits, if not set -->
+ <param name="machine_type" value="1"/><!-- 1 inorder; 0 OOO-->
+ <!-- inorder/OoO -->
+ <param name="number_hardware_threads" value="32"/>
+ <!-- number_instruction_fetch_ports(icache ports) is always 1 in single-thread processor,
+ it only may be more than one in SMT processors. BTB ports always equals to fetch ports since
+ branch information in consective branch instructions in the same fetch group can be read out from BTB once.-->
+ <param name="fetch_width" value="1"/>
+ <!-- fetch_width determins the size of cachelines of L1 cache block -->
+ <param name="number_instruction_fetch_ports" value="1"/>
+ <param name="decode_width" value="1"/>
+ <!-- decode_width determins the number of ports of the
+ renaming table (both RAM and CAM) scheme -->
+ <param name="issue_width" value="2"/>
+ <!-- issue_width determins the number of ports of Issue window and other logic
+ as in the complexity effective proccessors paper; issue_width==dispatch_width -->
+ <param name="commit_width" value="2"/>
+ <!-- commit_width determins the number of ports of register files -->
+ <param name="fp_issue_width" value="1"/>
+ <param name="prediction_width" value="0"/>
+ <!-- number of branch instructions can be predicted simultannouesl-->
+ <!-- Current version of McPAT does not distinguish int and floating point pipelines
+ Theses parameters are reserved for future use.-->
+ <param name="pipelines_per_core" value="1,1"/>
+ <!--integer_pipeline and floating_pipelines, if the floating_pipelines is 0, then the pipeline is shared-->
+ <param name="pipeline_depth" value="8,8"/>
+ <!-- pipeline depth of int and fp, if pipeline is shared, the second number is the average cycles of fp ops -->
+ <!-- issue and exe unit-->
+ <param name="ALU_per_core" value="32"/>
+ <!-- contains an adder, a shifter, and a logical unit -->
+ <param name="MUL_per_core" value="4"/>
+ <!-- For MUL and Div -->
+ <param name="FPU_per_core" value="32"/>
+ <!-- buffer between IF and ID stage -->
+ <param name="instruction_buffer_size" value="1"/>
+ <!-- buffer between ID and sche/exe stage -->
+ <param name="decoded_stream_buffer_size" value="1"/>
+ <param name="instruction_window_scheme" value="0"/><!-- 0 PHYREG based, 1 RSBASED-->
+ <!-- McPAT support 2 types of OoO cores, RS based and physical reg based-->
+ <param name="instruction_window_size" value="1"/>
+ <param name="fp_instruction_window_size" value="1"/>
+ <!-- the instruction issue Q as in Alpha 21264; The RS as in Intel P6 -->
+ <param name="ROB_size" value="0"/>
+ <!-- each in-flight instruction has an entry in ROB -->
+ <!-- registers -->
+ <!-- SM parameters Added by Syed Gilani -->
+ <param name="rf_banks" value="32"/>
+ <param name="simd_width" value="32"/>
+ <param name="collector_units" value="32"/>
+ <param name="core_clock_ratio" value="2"/>
+ <param name="warp_size" value="32"/>
+
+ <param name="archi_Regs_IRF_size" value="65536"/>
+ <param name="archi_Regs_FRF_size" value="32"/>
+ <!-- if OoO processor, phy_reg number is needed for renaming logic,
+ renaming logic is for both integer and floating point insts. -->
+ <param name="phy_Regs_IRF_size" value="32"/>
+ <param name="phy_Regs_FRF_size" value="32"/>
+ <!-- rename logic -->
+ <param name="rename_scheme" value="0"/>
+ <!-- can be RAM based(0) or CAM based(1) rename scheme
+ RAM-based scheme will have free list, status table;
+ CAM-based scheme have the valid bit in the data field of the CAM
+ both RAM and CAM need RAM-based checkpoint table, checkpoint_depth=# of in_flight instructions;
+ Detailed RAT Implementation see TR -->
+ <param name="register_windows_size" value="0"/>
+ <!-- how many windows in the windowed register file, sun processors;
+ no register windowing is used when this number is 0 -->
+ <!-- In OoO cores, loads and stores can be issued whether inorder(Pentium Pro) or (OoO)out-of-order(Alpha),
+ They will always try to exeute out-of-order though. -->
+ <param name="LSU_order" value="inorder"/>
+ <param name="store_buffer_size" value="32"/>
+ <!-- By default, in-order cores do not have load buffers -->
+ <param name="load_buffer_size" value="32"/>
+ <!-- number of ports refer to sustainable concurrent memory accesses -->
+ <param name="memory_ports" value="2"/>
+ <!-- max_allowed_in_flight_memo_instructions determins the # of ports of load and store buffer
+ as well as the ports of Dcache which is connected to LSU -->
+ <!-- dual-pumped Dcache can be used to save the extra read/write ports -->
+ <param name="RAS_size" value="1"/>
+ <!-- general stats, defines simulation periods;require total, idle, and busy cycles for senity check -->
+ <!-- please note: if target architecture is X86, then all the instrucions refer to (fused) micro-ops -->
+ <stat name="total_instructions" value="total_instructions_match_mcpat"/>
+ <stat name="int_instructions" value="int_instruction_match_mcpat"/>
+ <stat name="fp_instructions" value="flt_instruction_match_mcpat"/>
+ <stat name="branch_instructions" value="branch_instruction_match_mcpat"/>
+ <stat name="branch_mispredictions" value="0"/>
+ <stat name="load_instructions" value="load_instruction_match_mcpat"/>
+ <stat name="store_instructions" value="store_instruction_match_mcpat"/>
+ <stat name="committed_instructions" value="total_instructions_match_mcpat"/>
+ <stat name="committed_int_instructions" value="int_instruction_match_mcpat"/>
+ <stat name="committed_fp_instructions" value="flt_instruction_match_mcpat"/>
+ <stat name="pipeline_duty_cycle" value="0.6"/><!--<=1, runtime_ipc/peak_ipc; averaged for all cores if homogenous -->
+ <!-- the following cycle stats are used for heterogeneouse cores only,
+ please ignore them if homogeneouse cores -->
+ <stat name="total_cycles" value="total_cycles_match_mcpat"/>
+ <stat name="idle_cycles" value="idle_cycles_match_mcpat"/>
+ <stat name="busy_cycles" value="busy_cycles_match_mcpat"/>
+ <!-- instruction buffer stats -->
+ <!-- ROB stats, both RS and Phy based OoOs have ROB
+ performance simulator should capture the difference on accesses,
+ otherwise, McPAT has to guess based on number of commited instructions. -->
+ <stat name="ROB_reads" value="263886"/>
+ <stat name="ROB_writes" value="263886"/>
+ <!-- RAT accesses -->
+ <stat name="rename_accesses" value="263886"/>
+ <stat name="fp_rename_accesses" value="263886"/>
+ <!-- decode and rename stage use this, should be total ic - nop -->
+ <!-- Inst window stats -->
+ <stat name="inst_window_reads" value="263886"/>
+ <stat name="inst_window_writes" value="263886"/>
+ <stat name="inst_window_wakeup_accesses" value="263886"/>
+ <stat name="fp_inst_window_reads" value="263886"/>
+ <stat name="fp_inst_window_writes" value="263886"/>
+ <stat name="fp_inst_window_wakeup_accesses" value="263886"/>
+ <!-- RF accesses -->
+ <stat name="int_regfile_reads" value="int_register_read_access_match_mcpat"/>
+ <stat name="float_regfile_reads" value="int_register_write_access_match_mcpat"/>
+ <stat name="int_regfile_writes" value="float_register_read_access_match_mcpat"/>
+ <stat name="float_regfile_writes" value="float_register_write_access_match_mcpat"/>
+
+ <!-- The following stat is for operand collector power - Added by Syed -->
+ <stat name="non_rf_operands" value="0"/>
+
+ <!-- accesses to the working reg -->
+ <stat name="function_calls" value="0"/>
+ <stat name="context_switches" value="0"/> <!--not used in the McPAT -->
+ <!-- Number of Windowes switches (number of function calls and returns)-->
+ <!-- Alu stats by default, the processor has one FPU that includes the divider and
+ multiplier. The fpu accesses should include accesses to multiplier and divider -->
+ <stat name="ialu_accesses" value="ialu_accesses_match_mcpat"/>
+ <stat name="fpu_accesses" value="fpu_accesses_match_mcpat"/>
+ <stat name="mul_accesses" value="mul_accesses_match_mcpat"/>
+ <stat name="cdb_alu_accesses" value="0"/>
+ <stat name="cdb_mul_accesses" value="0"/>
+ <stat name="cdb_fpu_accesses" value="0"/>
+ <!-- multiple cycle accesses should be counted multiple times,
+ otherwise, McPAT can use internal counter for different floating point instructions
+ to get final accesses. But that needs detailed info for floating point inst mix -->
+ <!-- currently the performance simulator should
+ make sure all the numbers are final numbers,
+ including the explicit read/write accesses,
+ and the implicite accesses such as replacements and etc.
+ Future versions of McPAT may be able to reason the implicite access
+ based on param and stats of last level cache
+ The same rule applies to all cache access stats too! -->
+ <!-- following is AF for max power computation.
+ Do not change them, unless you understand them-->
+ <stat name="IFU_duty_cycle" value="0.25"/>
+ <stat name="LSU_duty_cycle" value="0.25"/>
+ <stat name="MemManU_I_duty_cycle" value="1"/>
+ <stat name="MemManU_D_duty_cycle" value="0.25"/>
+ <stat name="ALU_duty_cycle" value="0.9"/>
+ <stat name="MUL_duty_cycle" value="0.5"/>
+ <stat name="FPU_duty_cycle" value="1"/><!-- FPU numbers are already average -->
+ <stat name="ALU_cdb_duty_cycle" value="0.9"/>
+ <stat name="MUL_cdb_duty_cycle" value="0.5"/>
+ <stat name="FPU_cdb_duty_cycle" value="15"/>
+ <component id="system.core0.predictor" name="PBT">
+ <!-- branch predictor; tournament predictor see Alpha implementation -->
+ <param name="local_predictor_size" value="10,3"/>
+ <param name="local_predictor_entries" value="1024"/>
+ <param name="global_predictor_entries" value="4096"/>
+ <param name="global_predictor_bits" value="2"/>
+ <param name="chooser_predictor_entries" value="4096"/>
+ <param name="chooser_predictor_bits" value="2"/>
+ <!-- These parameters can be combined like below in next version
+ <param name="load_predictor" value="10,3,1024"/>
+ <param name="global_predictor" value="4096,2"/>
+ <param name="predictor_chooser" value="4096,2"/>
+ -->
+ </component>
+ <component id="system.core0.itlb" name="itlb">
+ <param name="number_entries" value="1"/>
+ <stat name="total_accesses" value="0"/>
+ <stat name="total_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <!-- there is no write requests to itlb although writes happen to itlb after miss,
+ which is actually a replacement -->
+ </component>
+ <component id="system.core0.icache" name="icache">
+ <!-- there is no write requests to itlb although writes happen to it after miss,
+ which is actually a replacement -->
+ <param name="icache_config" value="16384,128,4,1,1,3,8,0"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <!-- cache_policy;//0 no write or write-though with non-write allocate;1 write-back with write-allocate -->
+ <param name="buffer_sizes" value="16, 16, 16,0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="total_instructions_match_mcpat"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.dtlb" name="dtlb">
+ <param name="number_entries" value="1"/>
+ <stat name="total_accesses" value="0"/>
+ <stat name="total_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.ccache" name="ccache">
+ <!-- all the buffer related are optional -->
+ <param name="ccache_config" value="16384,64,2,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="ccache_read_accesses_match_mcpat"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="ccache_read_misses_match_mcpat"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.tcache" name="tcache">
+ <!-- all the buffer related are optional -->
+ <param name="tcache_config" value="49152,128,8,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="tcache_read_accesses_match_mcpat"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="tcache_read_misses_match_mcpat"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <!--model the shared memory by mimicing dcache-->
+ <component id="system.core0.sharedmemory" name="sharedmemory">
+ <!-- all the buffer related are optional -->
+ <param name="sharedmemory_config" value="98304,16,1,16,1,3,16,0"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="sharedmemory_read_access_match_mcpat"/>
+ <stat name="write_accesses" value="sharedmemory_write_access_match_mcpat"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.dcache" name="dcache">
+ <!-- all the buffer related are optional -->
+ <param name="dcache_config" value="16384,32,4,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="dcache_read_access_match_mcpat"/>
+ <stat name="write_accesses" value="dcache_write_access_match_mcpat"/>
+ <stat name="read_misses" value="dcache_read_miss_match_mcpat"/>
+ <stat name="write_misses" value="dcache_write_miss_match_mcpat"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.BTB" name="BTB">
+ <!-- all the buffer related are optional -->
+ <param name="BTB_config" value="8192,4,2,1, 1,3"/>
+ <!-- the parameters are capacity,block_width,associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ </component>
+ </component>
+ <component id="system.L1Directory0" name="L1Directory0">
+ <param name="Directory_type" value="0"/>
+ <!--0 cam based shadowed tag. 1 directory cache -->
+ <param name="Dir_config" value="2048,1,0,1, 4, 4,8"/>
+ <!-- the parameters are capacity,block_width, associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ <param name="buffer_sizes" value="8, 8, 8, 8"/>
+ <!-- all the buffer related are optional -->
+ <param name="clockrate" value="1400"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw search ports -->
+ <param name="device_type" value="0"/>
+ <!-- altough there are multiple access types,
+ Performance simulator needs to cast them into reads or writes
+ e.g. the invalidates can be considered as writes -->
+ <stat name="read_accesses" value="800000"/>
+ <stat name="write_accesses" value="27276"/>
+ <stat name="read_misses" value="1632"/>
+ <stat name="write_misses" value="183"/>
+ <stat name="conflicts" value="20"/>
+ <stat name="duty_cycle" value="0.45"/>
+ </component>
+ <component id="system.L2Directory0" name="L2Directory0">
+ <param name="Directory_type" value="1"/>
+ <!--0 cam based shadowed tag. 1 directory cache -->
+ <param name="Dir_config" value="1048576,16,16,1,2, 100"/>
+ <!-- the parameters are capacity,block_width, associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ <param name="buffer_sizes" value="8, 8, 8, 8"/>
+ <!-- all the buffer related are optional -->
+ <param name="clockrate" value="1400"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw search ports -->
+ <param name="device_type" value="0"/>
+ <!-- altough there are multiple access types,
+ Performance simulator needs to cast them into reads or writes
+ e.g. the invalidates can be considered as writes -->
+ <stat name="read_accesses" value="0"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.45"/>
+ </component>
+ <component id="system.L20" name="L20">
+ <!-- all the buffer related are optional -->
+ <param name="L2_config" value="131072,128,16,1, 4,23, 64, 1"/>
+ <!-- consider 4-way bank interleaving for Niagara 1 -->
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <param name="clockrate" value="2962"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw ports -->
+ <param name="device_type" value="0"/>
+ <stat name="read_accesses" value="200000"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.5"/>
+ </component>
+
+<!--**********************************************************************-->
+<component id="system.L30" name="L30">
+ <param name="L3_config" value="1048576,64,16,1, 2,100, 64,1"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="clockrate" value="3500"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw ports -->
+ <param name="device_type" value="0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="58824"/>
+ <stat name="write_accesses" value="27276"/>
+ <stat name="read_misses" value="1632"/>
+ <stat name="write_misses" value="183"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.35"/>
+ </component>
+
+
+<!--**********************************************************************-->
+ <component id="system.NoC0" name="noc0">
+ <param name="clockrate" value="700"/>
+ <param name="type" value="1"/>
+ <!-- 1 NoC, O bus -->
+ <param name="horizontal_nodes" value="2"/>
+ <param name="vertical_nodes" value="1"/>
+ <param name="has_global_link" value="0"/>
+ <!-- 1 has global link, 0 does not have global link -->
+ <param name="link_throughput" value="1"/><!--w.r.t clock -->
+ <param name="link_latency" value="1"/><!--w.r.t clock -->
+ <!-- througput >= latency -->
+ <!-- Router architecture -->
+ <param name="input_ports" value="6"/>
+ <param name="output_ports" value="6"/>
+ <param name="virtual_channel_per_port" value="1"/>
+ <!-- input buffer; in classic routers only input ports need buffers -->
+ <param name="flit_bits" value="32"/>
+ <param name="input_buffer_entries_per_vc" value="1"/><!--VCs within the same ports share input buffers whose size is propotional to the number of VCs-->
+ <param name="chip_coverage" value="1"/>
+ <!-- When multiple NOC present, one NOC will cover part of the whole chip. chip_coverage <=1 -->
+ <stat name="total_accesses" value="0"/>
+ <!-- This is the number of total accesses within the whole network not for each router -->
+ <stat name="duty_cycle" value="0.6"/>
+ </component>
+<!--**********************************************************************-->
+<!--**********************************************************************-->
+
+ <component id="system.mem" name="mem">
+ <!-- Main memory property -->
+ <param name="mem_tech_node" value="23"/>
+ <param name="device_clock" value="200"/><!--MHz, this is clock rate of the actual memory device, not the FSB -->
+ <param name="peak_transfer_rate" value="3200"/><!--MB/S-->
+ <param name="internal_prefetch_of_DRAM_chip" value="4"/>
+ <!-- 2 for DDR, 4 for DDR2, 8 for DDR3...-->
+ <!-- the device clock, peak_transfer_rate, and the internal prefetch decide the DIMM property -->
+ <!-- above numbers can be easily found from Wikipedia -->
+ <param name="capacity_per_channel" value="4096"/> <!-- MB -->
+ <!-- capacity_per_Dram_chip=capacity_per_channel/number_of_dimms/number_ranks/Dram_chips_per_rank
+ Current McPAT assumes single DIMMs are used.-->
+ <param name="number_ranks" value="2"/>
+ <param name="num_banks_of_DRAM_chip" value="6"/>
+ <param name="Block_width_of_DRAM_chip" value="64"/> <!-- B -->
+ <param name="output_width_of_DRAM_chip" value="8"/>
+ <!--number of Dram_chips_per_rank=" 72/output_width_of_DRAM_chip-->
+ <!--number of Dram_chips_per_rank=" 72/output_width_of_DRAM_chip-->
+ <param name="page_size_of_DRAM_chip" value="8"/> <!-- 8 or 16 -->
+ <param name="burstlength_of_DRAM_chip" value="8"/>
+ <stat name="memory_accesses" value="1052"/>
+ <stat name="memory_reads" value="1052"/>
+ <stat name="memory_writes" value="1052"/>
+ </component>
+ <component id="system.mc" name="mc">
+ <!-- Memeory controllers are for DDR(2,3...) DIMMs -->
+ <!-- current version of McPAT uses published values for base parameters of memory controller
+ improvments on MC will be added in later versions. -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="mc_clock" value="1848"/><!--DIMM IO bus clock rate MHz DDR2-400 for Niagara 1-->
+ <param name="peak_transfer_rate" value="29568"/><!--MB/S Syed: GTX 470 has 177.4GB/s mem transfer rate with 6 MCs -->
+ <param name="block_size" value="64"/><!--B-->
+ <param name="number_mcs" value="6"/><!-- 6 GDDR5 memory controllers -->
+ <!-- current McPAT only supports homogeneous memory controllers -->
+ <param name="memory_channels_per_mc" value="2"/>
+ <param name="number_ranks" value="1"/>
+ <param name="withPHY" value="0"/>
+ <!-- # of ranks of each channel-->
+ <param name="req_window_size_per_channel" value="16"/>
+ <param name="IO_buffer_size_per_channel" value="16"/>
+ <param name="databus_width" value="32"/>
+ <param name="addressbus_width" value="32"/>
+ <param name="PRT_entries" value="32"/>
+ <!-- # of empirical DRAM model parameter -->
+ <param name="dram_cmd_coeff" value="0"/>
+ <param name="dram_act_coeff" value="0"/>
+ <param name="dram_nop_coeff" value="0"/>
+ <param name="dram_activity_coeff" value="0"/>
+ <param name="dram_pre_coeff" value="3.8475e-8f"/>
+ <param name="dram_rd_coeff" value="7.74707143e-8f"/>
+ <param name="dram_wr_coeff" value="3.54664286e-8f"/>
+ <param name="dram_req_coeff" value="0"/>
+ <param name="dram_const_coeff" value="0"/>
+
+ <!-- McPAT will add the control bus width to the addressbus width automatically -->
+ <stat name="memory_accesses" value="memory_accesses_match_mcpat"/>
+ <stat name="memory_reads" value="memory_reads_match_mcpat"/>
+ <stat name="memory_writes" value="memory_writes_match_mcpat"/>
+ <!-- McPAT does not track individual mc, instead, it takes the total accesses and calculate
+ the average power per MC or per channel. This is sufficent for most application.
+ Further trackdown can be easily added in later versions. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.niu" name="niu">
+ <!-- On chip 10Gb Ethernet NIC, including XAUI Phy and MAC controller -->
+ <!-- For a minimum IP packet size of 84B at 10Gb/s, a new packet arrives every 67.2ns.
+ the low bound of clock rate of a 10Gb MAC is 150Mhz -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="clockrate" value="350"/>
+ <param name="number_units" value="0"/> <!-- unlike PCIe and memory controllers, each Ethernet controller only have one port -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- ratio of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual nic, instead, it takes the total accesses and calculate
+ the average power per nic or per channel. This is sufficent for most application. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.pcie" name="pcie">
+ <!-- On chip PCIe controller, including Phy-->
+ <!-- For a minimum PCIe packet size of 84B at 8Gb/s per lane (PCIe 3.0), a new packet arrives every 84ns.
+ the low bound of clock rate of a PCIe per lane logic is 120Mhz -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="withPHY" value="1"/>
+ <param name="clockrate" value="350"/>
+ <param name="number_units" value="0"/>
+ <param name="num_channels" value="8"/> <!-- 2 ,4 ,8 ,16 ,32 -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- Percentage of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual pcie controllers, instead, it takes the total accesses and calculate
+ the average power per pcie controller or per channel. This is sufficent for most application. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.flashc" name="flashc">
+ <param name="number_flashcs" value="0"/>
+ <param name="type" value="1"/> <!-- 1: low power; 0 high performance -->
+ <param name="withPHY" value="1"/>
+ <param name="peak_transfer_rate" value="200"/><!--Per controller sustainable reak rate MB/S -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- Percentage of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual flash controller, instead, it takes the total accesses and calculate
+ the average power per fc or per channel. This is sufficent for most application -->
+ </component>
+<!--**********************************************************************-->
+
+ </component>
+</component>
\ No newline at end of file diff --git a/configs/tested-cfgs/SM7_GV100/accelwattch_ptx_sim_alt.xml b/configs/tested-cfgs/SM7_GV100/accelwattch_ptx_sim_alt.xml new file mode 100644 index 0000000..0c6f211 --- /dev/null +++ b/configs/tested-cfgs/SM7_GV100/accelwattch_ptx_sim_alt.xml @@ -0,0 +1,623 @@ +<?xml version="1.0" ?>
+<component id="root" name="root">
+ <component id="system" name="system">
+<!-- Copyright (c) 2018-2021, Vijay Kandiah, Junrui Pan, Mahmoud Khairy, Scott Peverelle, Timothy Rogers, Tor M. Aamodt, Nikos Hardavellas
+Northwestern University, Purdue University, The University of British Columbia
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+ list of conditions and the following disclaimer;
+2. Redistributions in binary form must reproduce the above copyright notice,
+ this list of conditions and the following disclaimer in the documentation
+ and/or other materials provided with the distribution;
+3. Neither the names of Northwestern University, Purdue University,
+ The University of British Columbia nor the names of their contributors
+ may be used to endorse or promote products derived from this software
+ without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE. -->
+
+ <!-- DYNAMIC Power Activity Factors -->
+
+ <param name="TOT_INST" value="2" /> <!--Instruction Buffer Activity Factor(Total Warp Instructions)-->
+ <param name="FP_INT" value="4.57" /> <!--Scheduler Activity Factor(Non-Memory Warp Instructons)-->
+ <param name="IC_H" value="8.827711067" /> <!--Instruction Cache Hits -->
+ <param name="IC_M" value="8.827711067" /> <!--Instruction Cache Misses-->
+ <param name="DC_RH" value="8.545163843" /> <!--L1 Data Cache Read Hits-->
+ <param name="DC_RM" value="8.545163843" /> <!--L1 Data Cache Read Misses-->
+ <param name="DC_WH" value="8.545163843" /> <!--L1 Data Cache Write Hits-->
+ <param name="DC_WM" value="8.545163843" /> <!--L1 Data Cache Write Misses-->
+ <param name="CC_H" value="0.65916315" /> <!--Constant Cache Hits-->
+ <param name="CC_M" value="0.73418985" /> <!--Constant Cache Misses-->
+ <param name="SHRD_ACC" value="1.763326296" /> <!--Shared Memory Accesses-->
+ <param name="REG_RD" value="0.126123699" /> <!--Register Reads-->
+ <param name="REG_WR" value="0.126123699" /> <!--Register Writes-->
+ <param name="INT_ACC" value="2.851008299" /> <!--Integer ALU Accesses-->
+ <param name="FP_ACC" value="0.723984758" /> <!--FPU Accesses-->
+ <param name="DP_ACC" value="0.723979435" /> <!--Integer Accesses-->
+ <param name="INT_MUL_ACC" value="0.119939033" /> <!--Integer MUL Accesses-->
+ <param name="FP_MUL_ACC" value="0.147036868" /> <!--FP MUL Accesses-->
+ <param name="FP_SQRT_ACC" value="1.234782354" /> <!--SFU: sqrt Accesses-->
+ <param name="FP_LG_ACC" value="0.594393568" /> <!--SFU: log Accesses-->
+ <param name="FP_SIN_ACC" value="0.147037283" /> <!--SFU: sin/cos Accesses-->
+ <param name="FP_EXP_ACC" value="0.920933945" /> <!--SFU: exponent Accesses-->
+ <param name="DP_MUL_ACC" value="0.18425577" /> <!--DP MUL Accesses-->
+ <param name="TENSOR_ACC" value="1.656666667" /> <!--Tensor Accesses-->
+ <param name="TEX_ACC" value="0.147036868" /> <!--Texture Unit Accesses-->
+ <param name="MEM_RD" value="0.03058335" /> <!--DRAM Reads-->
+ <param name="MEM_WR" value="0.03058335" /> <!--DRAM Writes-->
+ <param name="MEM_PRE" value="0.03058335" /> <!--DRAM Precharge-->
+ <param name="L2_RH" value="2.032318132" /> <!--L2 Data Cache Read Hits-->
+ <param name="L2_RM" value="2.032318132" /> <!--L2 Data Cache Read Misses-->
+ <param name="L2_WH" value="2.032318132" /> <!--L2 Data Cache Write Hits-->
+ <param name="L2_WM" value="2.032318132" /> <!--L2 Data Cache Write Misses-->
+ <param name="NOC_A" value="70.15055938" /> <!--Interconnect Network Accesses-->
+ <param name="PIPE_A" value="0.01" /> <!--Pipeline Accesses-->
+
+ <!-- PTX Model specific DYNAMIC Power Activity Factors-->
+
+ <param name="TC_H" value="0.1" /> <!--Texture Cache Hits-->
+ <param name="TC_M" value="0.1" /> <!--Texture Cache Misses-->
+ <param name="INT_MUL24_ACC" value="0.235290825" /> <!--Integer 24-bit MUL Accesses-->
+ <param name="INT_MUL32_ACC" value="0.250616643" /> <!--Integer 32-bit MUL Accesses-->
+ <param name="INT_DIV_ACC" value="5.327155759" /> <!--Integer DIV Accesses-->
+ <param name="FP_DIV_ACC" value="4.992722677" /> <!--FP DIV Accesses-->
+ <param name="DP_DIV_ACC" value="26.46013946" /> <!--DP DIV Accesses-->
+
+ <!-- STATIC & CONSTANT Power -->
+
+ <param name="constant_power" value="32.32522272" /> <!--Constant power-->
+ <param name="idle_core_power" value="0.28279166"/> <!--Idle SM power-->
+
+ <param name="static_cat1_flane" value="15.29035866"/> <!--INT (ADD+MUL) First Lane Activation Power-->
+ <param name="static_cat1_addlane" value="0.586233603"/> <!--INT (ADD+MUL) Additional Lane Activation Power-->
+
+ <param name="static_cat2_flane" value="18.6179906"/> <!--INT+FP First Lane Activation Power-->
+ <param name="static_cat2_addlane" value="0.645228013"/> <!--INT+FP Additional Lane Activation Power-->
+
+ <param name="static_cat3_flane" value="19.10017723"/> <!--INT+FP+DP First Lane Activation Power-->
+ <param name="static_cat3_addlane" value="0.726863055"/> <!--INT+FP+DP Additional Lane Activation Power-->
+
+ <param name="static_cat4_flane" value="18.55029744"/> <!--INT+FP+SFU First Lane Activation Power-->
+ <param name="static_cat4_addlane" value="0.6099397"/> <!--INT+FP+SFU Additional Lane Activation Power-->
+
+ <param name="static_cat5_flane" value="14.74826681"/> <!--INT+FP+TEX First Lane Activation Power-->
+ <param name="static_cat5_addlane" value="0.514367937"/> <!--INT+FP+TEX Additional Lane Activation Power-->
+
+ <param name="static_cat6_flane" value="48.94875596"/> <!--INT+FP+TENSOR First Lane Activation Power-->
+ <param name="static_cat6_addlane" value="0.0"/> <!--INT+FP+TENSOR Additional Lane Activation Power-->
+
+ <param name="static_light_flane" value="1.965373811"/> <!--LIGHT_SM First Lane Activation Power-->
+ <param name="static_light_addlane" value="0.003966868"/> <!--LIGHT_SM Additional Lane Activation Power-->
+
+ <param name="static_intadd_flane" value="19.70468506"/> <!--INT ADD First Lane Activation Power-->
+ <param name="static_intadd_addlane" value="0.388578623"/> <!--INT ADD Additional Lane Activation Power-->
+
+ <param name="static_intmul_flane" value="16.64811823"/> <!--INT MUL First Lane Activation Power-->
+ <param name="static_intmul_addlane" value="0.281803166"/> <!--INT MUL Additional Lane Activation Power-->
+
+ <param name="static_geomean_flane" value="17.21745077"/> <!--GEOMEAN First Lane Activation Power-->
+ <param name="static_geomean_addlane" value="0.650630555"/> <!--GEOMEAN Additional Lane Activation Power-->
+ <param name="static_shared_flane" value="31.40965691"/> <!--Shared Memory First Lane Activation Power-->
+ <param name="static_l1_flane" value="34.79491352"/> <!--L1D First Lane Activation Power-->
+ <param name="static_l2_flane" value="17.30654755"/> <!--L2D ADD+MUL First Lane Activation Power-->
+
+
+ <!-- LEAVE BELOW UNCHANGED! These are parameters used for initial GPUWattch model left untouched -->
+
+
+ <param name="GPU_Architecture" value="1"/><!-- 0-G80; 1-Fermi; others not supported -->
+ <param name="architecture" value="1"/> <!-- fermi:1 quadro:2 other: undefined-->
+ <param name="number_of_L1Directories" value="0"/>
+ <param name="number_of_L2Directories" value="0"/>
+ <param name="number_of_L2s" value="1"/> <!-- This number means how many L2 clusters in each cluster there can be multiple banks/ports -->
+ <param name="number_of_L3s" value="0"/> <!-- This number means how many L3 clusters -->
+ <param name="number_of_NoCs" value="1"/>
+ <param name="homogeneous_cores" value="1"/><!--1 means homo -->
+ <param name="homogeneous_L2s" value="1"/>
+ <param name="homogeneous_L1Directorys" value="1"/>
+ <param name="homogeneous_L2Directorys" value="1"/>
+ <param name="homogeneous_L3s" value="1"/>
+ <param name="homogeneous_ccs" value="1"/><!--cache coherece hardware -->
+ <param name="homogeneous_NoCs" value="1"/>
+ <param name="core_tech_node" value="23"/><!-- nm -->
+
+ <param name="number_cache_levels" value="2"/>
+ <param name="interconnect_projection_type" value="0"/><!--0: agressive wire technology; 1: conservative wire technology -->
+ <param name="device_type" value="0"/><!--0: HP(High Performance Type); 1: LSTP(Low standby power) 2: LOP (Low Operating Power) -->
+ <param name="longer_channel_device" value="1"/><!-- 0 no use; 1 use when possible -->
+ <param name="machine_bits" value="32"/>
+ <param name="virtual_address_width" value="32"/>
+ <param name="physical_address_width" value="32"/>
+ <param name="virtual_memory_page_size" value="4096"/>
+
+ <param name="IDLE_CORE_N" value="1"/>
+ <stat name="num_idle_cores" value="0"/><!-- Average Number of idle cores during this period -->
+ <stat name="total_cycles" value="total_cycles_match_mcpat"/>
+ <stat name="idle_cycles" value="idle_cycles_match_mcpat"/>
+ <stat name="busy_cycles" value="busy_cycles_match_mcpat"/>
+ <!--This page size(B) is complete different from the page size in Main memo secction. this page size is the size of
+ virtual memory from OS/Archi perspective; the page size in Main memo secction is the actuall physical line in a DRAM bank -->
+ <!-- *********************** cores ******************* -->
+ <component id="system.core0" name="core0">
+ <!-- Core property -->
+ <param name="clock_rate" value="700"/>
+ <param name="instruction_length" value="32"/>
+ <param name="opcode_width" value="9"/>
+ <!-- address width determins the tag_width in Cache, LSQ and buffers in cache controller
+ default value is machine_bits, if not set -->
+ <param name="machine_type" value="1"/><!-- 1 inorder; 0 OOO-->
+ <!-- inorder/OoO -->
+ <param name="number_hardware_threads" value="32"/>
+ <!-- number_instruction_fetch_ports(icache ports) is always 1 in single-thread processor,
+ it only may be more than one in SMT processors. BTB ports always equals to fetch ports since
+ branch information in consective branch instructions in the same fetch group can be read out from BTB once.-->
+ <param name="fetch_width" value="1"/>
+ <!-- fetch_width determins the size of cachelines of L1 cache block -->
+ <param name="number_instruction_fetch_ports" value="1"/>
+ <param name="decode_width" value="1"/>
+ <!-- decode_width determins the number of ports of the
+ renaming table (both RAM and CAM) scheme -->
+ <param name="issue_width" value="2"/>
+ <!-- issue_width determins the number of ports of Issue window and other logic
+ as in the complexity effective proccessors paper; issue_width==dispatch_width -->
+ <param name="commit_width" value="2"/>
+ <!-- commit_width determins the number of ports of register files -->
+ <param name="fp_issue_width" value="1"/>
+ <param name="prediction_width" value="0"/>
+ <!-- number of branch instructions can be predicted simultannouesl-->
+ <!-- Current version of McPAT does not distinguish int and floating point pipelines
+ Theses parameters are reserved for future use.-->
+ <param name="pipelines_per_core" value="1,1"/>
+ <!--integer_pipeline and floating_pipelines, if the floating_pipelines is 0, then the pipeline is shared-->
+ <param name="pipeline_depth" value="8,8"/>
+ <!-- pipeline depth of int and fp, if pipeline is shared, the second number is the average cycles of fp ops -->
+ <!-- issue and exe unit-->
+ <param name="ALU_per_core" value="32"/>
+ <!-- contains an adder, a shifter, and a logical unit -->
+ <param name="MUL_per_core" value="4"/>
+ <!-- For MUL and Div -->
+ <param name="FPU_per_core" value="32"/>
+ <!-- buffer between IF and ID stage -->
+ <param name="instruction_buffer_size" value="1"/>
+ <!-- buffer between ID and sche/exe stage -->
+ <param name="decoded_stream_buffer_size" value="1"/>
+ <param name="instruction_window_scheme" value="0"/><!-- 0 PHYREG based, 1 RSBASED-->
+ <!-- McPAT support 2 types of OoO cores, RS based and physical reg based-->
+ <param name="instruction_window_size" value="1"/>
+ <param name="fp_instruction_window_size" value="1"/>
+ <!-- the instruction issue Q as in Alpha 21264; The RS as in Intel P6 -->
+ <param name="ROB_size" value="0"/>
+ <!-- each in-flight instruction has an entry in ROB -->
+ <!-- registers -->
+ <!-- SM parameters Added by Syed Gilani -->
+ <param name="rf_banks" value="32"/>
+ <param name="simd_width" value="32"/>
+ <param name="collector_units" value="32"/>
+ <param name="core_clock_ratio" value="2"/>
+ <param name="warp_size" value="32"/>
+
+ <param name="archi_Regs_IRF_size" value="65536"/>
+ <param name="archi_Regs_FRF_size" value="32"/>
+ <!-- if OoO processor, phy_reg number is needed for renaming logic,
+ renaming logic is for both integer and floating point insts. -->
+ <param name="phy_Regs_IRF_size" value="32"/>
+ <param name="phy_Regs_FRF_size" value="32"/>
+ <!-- rename logic -->
+ <param name="rename_scheme" value="0"/>
+ <!-- can be RAM based(0) or CAM based(1) rename scheme
+ RAM-based scheme will have free list, status table;
+ CAM-based scheme have the valid bit in the data field of the CAM
+ both RAM and CAM need RAM-based checkpoint table, checkpoint_depth=# of in_flight instructions;
+ Detailed RAT Implementation see TR -->
+ <param name="register_windows_size" value="0"/>
+ <!-- how many windows in the windowed register file, sun processors;
+ no register windowing is used when this number is 0 -->
+ <!-- In OoO cores, loads and stores can be issued whether inorder(Pentium Pro) or (OoO)out-of-order(Alpha),
+ They will always try to exeute out-of-order though. -->
+ <param name="LSU_order" value="inorder"/>
+ <param name="store_buffer_size" value="32"/>
+ <!-- By default, in-order cores do not have load buffers -->
+ <param name="load_buffer_size" value="32"/>
+ <!-- number of ports refer to sustainable concurrent memory accesses -->
+ <param name="memory_ports" value="2"/>
+ <!-- max_allowed_in_flight_memo_instructions determins the # of ports of load and store buffer
+ as well as the ports of Dcache which is connected to LSU -->
+ <!-- dual-pumped Dcache can be used to save the extra read/write ports -->
+ <param name="RAS_size" value="1"/>
+ <!-- general stats, defines simulation periods;require total, idle, and busy cycles for senity check -->
+ <!-- please note: if target architecture is X86, then all the instrucions refer to (fused) micro-ops -->
+ <stat name="total_instructions" value="total_instructions_match_mcpat"/>
+ <stat name="int_instructions" value="int_instruction_match_mcpat"/>
+ <stat name="fp_instructions" value="flt_instruction_match_mcpat"/>
+ <stat name="branch_instructions" value="branch_instruction_match_mcpat"/>
+ <stat name="branch_mispredictions" value="0"/>
+ <stat name="load_instructions" value="load_instruction_match_mcpat"/>
+ <stat name="store_instructions" value="store_instruction_match_mcpat"/>
+ <stat name="committed_instructions" value="total_instructions_match_mcpat"/>
+ <stat name="committed_int_instructions" value="int_instruction_match_mcpat"/>
+ <stat name="committed_fp_instructions" value="flt_instruction_match_mcpat"/>
+ <stat name="pipeline_duty_cycle" value="0.6"/><!--<=1, runtime_ipc/peak_ipc; averaged for all cores if homogenous -->
+ <!-- the following cycle stats are used for heterogeneouse cores only,
+ please ignore them if homogeneouse cores -->
+ <stat name="total_cycles" value="total_cycles_match_mcpat"/>
+ <stat name="idle_cycles" value="idle_cycles_match_mcpat"/>
+ <stat name="busy_cycles" value="busy_cycles_match_mcpat"/>
+ <!-- instruction buffer stats -->
+ <!-- ROB stats, both RS and Phy based OoOs have ROB
+ performance simulator should capture the difference on accesses,
+ otherwise, McPAT has to guess based on number of commited instructions. -->
+ <stat name="ROB_reads" value="263886"/>
+ <stat name="ROB_writes" value="263886"/>
+ <!-- RAT accesses -->
+ <stat name="rename_accesses" value="263886"/>
+ <stat name="fp_rename_accesses" value="263886"/>
+ <!-- decode and rename stage use this, should be total ic - nop -->
+ <!-- Inst window stats -->
+ <stat name="inst_window_reads" value="263886"/>
+ <stat name="inst_window_writes" value="263886"/>
+ <stat name="inst_window_wakeup_accesses" value="263886"/>
+ <stat name="fp_inst_window_reads" value="263886"/>
+ <stat name="fp_inst_window_writes" value="263886"/>
+ <stat name="fp_inst_window_wakeup_accesses" value="263886"/>
+ <!-- RF accesses -->
+ <stat name="int_regfile_reads" value="int_register_read_access_match_mcpat"/>
+ <stat name="float_regfile_reads" value="int_register_write_access_match_mcpat"/>
+ <stat name="int_regfile_writes" value="float_register_read_access_match_mcpat"/>
+ <stat name="float_regfile_writes" value="float_register_write_access_match_mcpat"/>
+
+ <!-- The following stat is for operand collector power - Added by Syed -->
+ <stat name="non_rf_operands" value="0"/>
+
+ <!-- accesses to the working reg -->
+ <stat name="function_calls" value="0"/>
+ <stat name="context_switches" value="0"/> <!--not used in the McPAT -->
+ <!-- Number of Windowes switches (number of function calls and returns)-->
+ <!-- Alu stats by default, the processor has one FPU that includes the divider and
+ multiplier. The fpu accesses should include accesses to multiplier and divider -->
+ <stat name="ialu_accesses" value="ialu_accesses_match_mcpat"/>
+ <stat name="fpu_accesses" value="fpu_accesses_match_mcpat"/>
+ <stat name="mul_accesses" value="mul_accesses_match_mcpat"/>
+ <stat name="cdb_alu_accesses" value="0"/>
+ <stat name="cdb_mul_accesses" value="0"/>
+ <stat name="cdb_fpu_accesses" value="0"/>
+ <!-- multiple cycle accesses should be counted multiple times,
+ otherwise, McPAT can use internal counter for different floating point instructions
+ to get final accesses. But that needs detailed info for floating point inst mix -->
+ <!-- currently the performance simulator should
+ make sure all the numbers are final numbers,
+ including the explicit read/write accesses,
+ and the implicite accesses such as replacements and etc.
+ Future versions of McPAT may be able to reason the implicite access
+ based on param and stats of last level cache
+ The same rule applies to all cache access stats too! -->
+ <!-- following is AF for max power computation.
+ Do not change them, unless you understand them-->
+ <stat name="IFU_duty_cycle" value="0.25"/>
+ <stat name="LSU_duty_cycle" value="0.25"/>
+ <stat name="MemManU_I_duty_cycle" value="1"/>
+ <stat name="MemManU_D_duty_cycle" value="0.25"/>
+ <stat name="ALU_duty_cycle" value="0.9"/>
+ <stat name="MUL_duty_cycle" value="0.5"/>
+ <stat name="FPU_duty_cycle" value="1"/><!-- FPU numbers are already average -->
+ <stat name="ALU_cdb_duty_cycle" value="0.9"/>
+ <stat name="MUL_cdb_duty_cycle" value="0.5"/>
+ <stat name="FPU_cdb_duty_cycle" value="15"/>
+ <component id="system.core0.predictor" name="PBT">
+ <!-- branch predictor; tournament predictor see Alpha implementation -->
+ <param name="local_predictor_size" value="10,3"/>
+ <param name="local_predictor_entries" value="1024"/>
+ <param name="global_predictor_entries" value="4096"/>
+ <param name="global_predictor_bits" value="2"/>
+ <param name="chooser_predictor_entries" value="4096"/>
+ <param name="chooser_predictor_bits" value="2"/>
+ <!-- These parameters can be combined like below in next version
+ <param name="load_predictor" value="10,3,1024"/>
+ <param name="global_predictor" value="4096,2"/>
+ <param name="predictor_chooser" value="4096,2"/>
+ -->
+ </component>
+ <component id="system.core0.itlb" name="itlb">
+ <param name="number_entries" value="1"/>
+ <stat name="total_accesses" value="0"/>
+ <stat name="total_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <!-- there is no write requests to itlb although writes happen to itlb after miss,
+ which is actually a replacement -->
+ </component>
+ <component id="system.core0.icache" name="icache">
+ <!-- there is no write requests to itlb although writes happen to it after miss,
+ which is actually a replacement -->
+ <param name="icache_config" value="16384,128,4,1,1,3,8,0"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <!-- cache_policy;//0 no write or write-though with non-write allocate;1 write-back with write-allocate -->
+ <param name="buffer_sizes" value="16, 16, 16,0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="total_instructions_match_mcpat"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.dtlb" name="dtlb">
+ <param name="number_entries" value="1"/>
+ <stat name="total_accesses" value="0"/>
+ <stat name="total_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.ccache" name="ccache">
+ <!-- all the buffer related are optional -->
+ <param name="ccache_config" value="16384,64,2,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="ccache_read_accesses_match_mcpat"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="ccache_read_misses_match_mcpat"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.tcache" name="tcache">
+ <!-- all the buffer related are optional -->
+ <param name="tcache_config" value="49152,128,8,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="tcache_read_accesses_match_mcpat"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="tcache_read_misses_match_mcpat"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <!--model the shared memory by mimicing dcache-->
+ <component id="system.core0.sharedmemory" name="sharedmemory">
+ <!-- all the buffer related are optional -->
+ <param name="sharedmemory_config" value="98304,16,1,16,1,3,16,0"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="sharedmemory_read_access_match_mcpat"/>
+ <stat name="write_accesses" value="sharedmemory_write_access_match_mcpat"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.dcache" name="dcache">
+ <!-- all the buffer related are optional -->
+ <param name="dcache_config" value="16384,32,4,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="dcache_read_access_match_mcpat"/>
+ <stat name="write_accesses" value="dcache_write_access_match_mcpat"/>
+ <stat name="read_misses" value="dcache_read_miss_match_mcpat"/>
+ <stat name="write_misses" value="dcache_write_miss_match_mcpat"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.BTB" name="BTB">
+ <!-- all the buffer related are optional -->
+ <param name="BTB_config" value="8192,4,2,1, 1,3"/>
+ <!-- the parameters are capacity,block_width,associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ </component>
+ </component>
+ <component id="system.L1Directory0" name="L1Directory0">
+ <param name="Directory_type" value="0"/>
+ <!--0 cam based shadowed tag. 1 directory cache -->
+ <param name="Dir_config" value="2048,1,0,1, 4, 4,8"/>
+ <!-- the parameters are capacity,block_width, associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ <param name="buffer_sizes" value="8, 8, 8, 8"/>
+ <!-- all the buffer related are optional -->
+ <param name="clockrate" value="1400"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw search ports -->
+ <param name="device_type" value="0"/>
+ <!-- altough there are multiple access types,
+ Performance simulator needs to cast them into reads or writes
+ e.g. the invalidates can be considered as writes -->
+ <stat name="read_accesses" value="800000"/>
+ <stat name="write_accesses" value="27276"/>
+ <stat name="read_misses" value="1632"/>
+ <stat name="write_misses" value="183"/>
+ <stat name="conflicts" value="20"/>
+ <stat name="duty_cycle" value="0.45"/>
+ </component>
+ <component id="system.L2Directory0" name="L2Directory0">
+ <param name="Directory_type" value="1"/>
+ <!--0 cam based shadowed tag. 1 directory cache -->
+ <param name="Dir_config" value="1048576,16,16,1,2, 100"/>
+ <!-- the parameters are capacity,block_width, associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ <param name="buffer_sizes" value="8, 8, 8, 8"/>
+ <!-- all the buffer related are optional -->
+ <param name="clockrate" value="1400"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw search ports -->
+ <param name="device_type" value="0"/>
+ <!-- altough there are multiple access types,
+ Performance simulator needs to cast them into reads or writes
+ e.g. the invalidates can be considered as writes -->
+ <stat name="read_accesses" value="0"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.45"/>
+ </component>
+ <component id="system.L20" name="L20">
+ <!-- all the buffer related are optional -->
+ <param name="L2_config" value="131072,128,16,1, 4,23, 64, 1"/>
+ <!-- consider 4-way bank interleaving for Niagara 1 -->
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <param name="clockrate" value="2962"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw ports -->
+ <param name="device_type" value="0"/>
+ <stat name="read_accesses" value="200000"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.5"/>
+ </component>
+
+<!--**********************************************************************-->
+<component id="system.L30" name="L30">
+ <param name="L3_config" value="1048576,64,16,1, 2,100, 64,1"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="clockrate" value="3500"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw ports -->
+ <param name="device_type" value="0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="58824"/>
+ <stat name="write_accesses" value="27276"/>
+ <stat name="read_misses" value="1632"/>
+ <stat name="write_misses" value="183"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.35"/>
+ </component>
+
+
+<!--**********************************************************************-->
+ <component id="system.NoC0" name="noc0">
+ <param name="clockrate" value="700"/>
+ <param name="type" value="1"/>
+ <!-- 1 NoC, O bus -->
+ <param name="horizontal_nodes" value="2"/>
+ <param name="vertical_nodes" value="1"/>
+ <param name="has_global_link" value="0"/>
+ <!-- 1 has global link, 0 does not have global link -->
+ <param name="link_throughput" value="1"/><!--w.r.t clock -->
+ <param name="link_latency" value="1"/><!--w.r.t clock -->
+ <!-- througput >= latency -->
+ <!-- Router architecture -->
+ <param name="input_ports" value="6"/>
+ <param name="output_ports" value="6"/>
+ <param name="virtual_channel_per_port" value="1"/>
+ <!-- input buffer; in classic routers only input ports need buffers -->
+ <param name="flit_bits" value="32"/>
+ <param name="input_buffer_entries_per_vc" value="1"/><!--VCs within the same ports share input buffers whose size is propotional to the number of VCs-->
+ <param name="chip_coverage" value="1"/>
+ <!-- When multiple NOC present, one NOC will cover part of the whole chip. chip_coverage <=1 -->
+ <stat name="total_accesses" value="0"/>
+ <!-- This is the number of total accesses within the whole network not for each router -->
+ <stat name="duty_cycle" value="0.6"/>
+ </component>
+<!--**********************************************************************-->
+<!--**********************************************************************-->
+
+ <component id="system.mem" name="mem">
+ <!-- Main memory property -->
+ <param name="mem_tech_node" value="23"/>
+ <param name="device_clock" value="200"/><!--MHz, this is clock rate of the actual memory device, not the FSB -->
+ <param name="peak_transfer_rate" value="3200"/><!--MB/S-->
+ <param name="internal_prefetch_of_DRAM_chip" value="4"/>
+ <!-- 2 for DDR, 4 for DDR2, 8 for DDR3...-->
+ <!-- the device clock, peak_transfer_rate, and the internal prefetch decide the DIMM property -->
+ <!-- above numbers can be easily found from Wikipedia -->
+ <param name="capacity_per_channel" value="4096"/> <!-- MB -->
+ <!-- capacity_per_Dram_chip=capacity_per_channel/number_of_dimms/number_ranks/Dram_chips_per_rank
+ Current McPAT assumes single DIMMs are used.-->
+ <param name="number_ranks" value="2"/>
+ <param name="num_banks_of_DRAM_chip" value="6"/>
+ <param name="Block_width_of_DRAM_chip" value="64"/> <!-- B -->
+ <param name="output_width_of_DRAM_chip" value="8"/>
+ <!--number of Dram_chips_per_rank=" 72/output_width_of_DRAM_chip-->
+ <!--number of Dram_chips_per_rank=" 72/output_width_of_DRAM_chip-->
+ <param name="page_size_of_DRAM_chip" value="8"/> <!-- 8 or 16 -->
+ <param name="burstlength_of_DRAM_chip" value="8"/>
+ <stat name="memory_accesses" value="1052"/>
+ <stat name="memory_reads" value="1052"/>
+ <stat name="memory_writes" value="1052"/>
+ </component>
+ <component id="system.mc" name="mc">
+ <!-- Memeory controllers are for DDR(2,3...) DIMMs -->
+ <!-- current version of McPAT uses published values for base parameters of memory controller
+ improvments on MC will be added in later versions. -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="mc_clock" value="1848"/><!--DIMM IO bus clock rate MHz DDR2-400 for Niagara 1-->
+ <param name="peak_transfer_rate" value="29568"/><!--MB/S Syed: GTX 470 has 177.4GB/s mem transfer rate with 6 MCs -->
+ <param name="block_size" value="64"/><!--B-->
+ <param name="number_mcs" value="6"/><!-- 6 GDDR5 memory controllers -->
+ <!-- current McPAT only supports homogeneous memory controllers -->
+ <param name="memory_channels_per_mc" value="2"/>
+ <param name="number_ranks" value="1"/>
+ <param name="withPHY" value="0"/>
+ <!-- # of ranks of each channel-->
+ <param name="req_window_size_per_channel" value="16"/>
+ <param name="IO_buffer_size_per_channel" value="16"/>
+ <param name="databus_width" value="32"/>
+ <param name="addressbus_width" value="32"/>
+ <param name="PRT_entries" value="32"/>
+ <!-- # of empirical DRAM model parameter -->
+ <param name="dram_cmd_coeff" value="0"/>
+ <param name="dram_act_coeff" value="0"/>
+ <param name="dram_nop_coeff" value="0"/>
+ <param name="dram_activity_coeff" value="0"/>
+ <param name="dram_pre_coeff" value="3.8475e-8f"/>
+ <param name="dram_rd_coeff" value="7.74707143e-8f"/>
+ <param name="dram_wr_coeff" value="3.54664286e-8f"/>
+ <param name="dram_req_coeff" value="0"/>
+ <param name="dram_const_coeff" value="0"/>
+
+ <!-- McPAT will add the control bus width to the addressbus width automatically -->
+ <stat name="memory_accesses" value="memory_accesses_match_mcpat"/>
+ <stat name="memory_reads" value="memory_reads_match_mcpat"/>
+ <stat name="memory_writes" value="memory_writes_match_mcpat"/>
+ <!-- McPAT does not track individual mc, instead, it takes the total accesses and calculate
+ the average power per MC or per channel. This is sufficent for most application.
+ Further trackdown can be easily added in later versions. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.niu" name="niu">
+ <!-- On chip 10Gb Ethernet NIC, including XAUI Phy and MAC controller -->
+ <!-- For a minimum IP packet size of 84B at 10Gb/s, a new packet arrives every 67.2ns.
+ the low bound of clock rate of a 10Gb MAC is 150Mhz -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="clockrate" value="350"/>
+ <param name="number_units" value="0"/> <!-- unlike PCIe and memory controllers, each Ethernet controller only have one port -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- ratio of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual nic, instead, it takes the total accesses and calculate
+ the average power per nic or per channel. This is sufficent for most application. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.pcie" name="pcie">
+ <!-- On chip PCIe controller, including Phy-->
+ <!-- For a minimum PCIe packet size of 84B at 8Gb/s per lane (PCIe 3.0), a new packet arrives every 84ns.
+ the low bound of clock rate of a PCIe per lane logic is 120Mhz -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="withPHY" value="1"/>
+ <param name="clockrate" value="350"/>
+ <param name="number_units" value="0"/>
+ <param name="num_channels" value="8"/> <!-- 2 ,4 ,8 ,16 ,32 -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- Percentage of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual pcie controllers, instead, it takes the total accesses and calculate
+ the average power per pcie controller or per channel. This is sufficent for most application. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.flashc" name="flashc">
+ <param name="number_flashcs" value="0"/>
+ <param name="type" value="1"/> <!-- 1: low power; 0 high performance -->
+ <param name="withPHY" value="1"/>
+ <param name="peak_transfer_rate" value="200"/><!--Per controller sustainable reak rate MB/S -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- Percentage of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual flash controller, instead, it takes the total accesses and calculate
+ the average power per fc or per channel. This is sufficent for most application -->
+ </component>
+<!--**********************************************************************-->
+
+ </component>
+</component>
\ No newline at end of file diff --git a/configs/tested-cfgs/SM7_GV100/accelwattch_sass_hw.xml b/configs/tested-cfgs/SM7_GV100/accelwattch_sass_hw.xml new file mode 100644 index 0000000..64f89d6 --- /dev/null +++ b/configs/tested-cfgs/SM7_GV100/accelwattch_sass_hw.xml @@ -0,0 +1,613 @@ +<?xml version="1.0" ?>
+<component id="root" name="root">
+ <component id="system" name="system">
+<!-- Copyright (c) 2018-2021, Vijay Kandiah, Junrui Pan, Mahmoud Khairy, Scott Peverelle, Timothy Rogers, Tor M. Aamodt, Nikos Hardavellas
+Northwestern University, Purdue University, The University of British Columbia
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+ list of conditions and the following disclaimer;
+2. Redistributions in binary form must reproduce the above copyright notice,
+ this list of conditions and the following disclaimer in the documentation
+ and/or other materials provided with the distribution;
+3. Neither the names of Northwestern University, Purdue University,
+ The University of British Columbia nor the names of their contributors
+ may be used to endorse or promote products derived from this software
+ without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE. -->
+
+ <!-- DYNAMIC Power Activity Factors -->
+
+ <param name="TOT_INST" value="100" /> <!--Instruction Buffer Activity Factor(Total Warp Instructions)-->
+ <param name="FP_INT" value="43.5416637" /> <!--Scheduler Activity Factor(Non-Memory Warp Instructons)-->
+ <param name="IC_H" value="0" /> <!--Instruction Cache Hits -->
+ <param name="IC_M" value="0" /> <!--Instruction Cache Misses-->
+ <param name="DC_RH" value="7.193244877" /> <!--L1 Data Cache Read Hits-->
+ <param name="DC_RM" value="8.01198819" /> <!--L1 Data Cache Read Misses-->
+ <param name="DC_WH" value="0.497094158" /> <!--L1 Data Cache Write Hits-->
+ <param name="DC_WM" value="12.9276971" /> <!--L1 Data Cache Write Misses-->
+ <param name="CC_H" value="9.322047" /> <!--Constant Cache Hits-->
+ <param name="CC_M" value="10.383093" /> <!--Constant Cache Misses-->
+ <param name="SHRD_ACC" value="4.146674883" /> <!--Shared Memory Accesses-->
+ <param name="REG_RD" value="0" /> <!--Register Reads-->
+ <param name="REG_WR" value="0" /> <!--Register Writes-->
+ <param name="INT_ACC" value="3.350646078" /> <!--Integer ALU Accesses-->
+ <param name="FP_ACC" value="0.041043127" /> <!--FPU Accesses-->
+ <param name="DP_ACC" value="0.077722905" /> <!--Integer Accesses-->
+ <param name="INT_MUL_ACC" value="0.147866461" /> <!--Integer MUL Accesses-->
+ <param name="FP_MUL_ACC" value="0.147864271" /> <!--FP MUL Accesses-->
+ <param name="FP_SQRT_ACC" value="0.48526506" /> <!--SFU: sqrt Accesses-->
+ <param name="FP_LG_ACC" value="0.147864519" /> <!--SFU: log Accesses-->
+ <param name="FP_SIN_ACC" value="0.376510619" /> <!--SFU: sin/cos Accesses-->
+ <param name="FP_EXP_ACC" value="0.147862451" /> <!--SFU: exponent Accesses-->
+ <param name="DP_MUL_ACC" value="0.147865348" /> <!--DP MUL Accesses-->
+ <param name="TENSOR_ACC" value="0.148632914" /> <!--Tensor Accesses-->
+ <param name="TEX_ACC" value="3.4530026" /> <!--Texture Unit Accesses-->
+ <param name="MEM_RD" value="0.020752854" /> <!--DRAM Reads-->
+ <param name="MEM_WR" value="0.025154975" /> <!--DRAM Writes-->
+ <param name="MEM_PRE" value="0.006917618" /> <!--DRAM Precharge-->
+ <param name="L2_RH" value="3.430442278" /> <!--L2 Data Cache Read Hits-->
+ <param name="L2_RM" value="6.514812193" /> <!--L2 Data Cache Read Misses-->
+ <param name="L2_WH" value="11.22265896" /> <!--L2 Data Cache Write Hits-->
+ <param name="L2_WM" value="3.32662057" /> <!--L2 Data Cache Write Misses-->
+ <param name="NOC_A" value="87.30828879" /> <!--Interconnect Network Accesses-->
+ <param name="PIPE_A" value="1.028" /> <!--Pipeline Accesses-->
+
+ <!-- STATIC & CONSTANT Power -->
+
+ <param name="constant_power" value="32.32522272" /> <!--Constant power-->
+ <param name="idle_core_power" value="0.28279166"/> <!--Idle SM power-->
+
+ <param name="static_cat1_flane" value="15.29035866"/> <!--INT (ADD+MUL) First Lane Activation Power-->
+ <param name="static_cat1_addlane" value="0.586233603"/> <!--INT (ADD+MUL) Additional Lane Activation Power-->
+
+ <param name="static_cat2_flane" value="18.6179906"/> <!--INT+FP First Lane Activation Power-->
+ <param name="static_cat2_addlane" value="0.645228013"/> <!--INT+FP Additional Lane Activation Power-->
+
+ <param name="static_cat3_flane" value="19.10017723"/> <!--INT+FP+DP First Lane Activation Power-->
+ <param name="static_cat3_addlane" value="0.726863055"/> <!--INT+FP+DP Additional Lane Activation Power-->
+
+ <param name="static_cat4_flane" value="18.55029744"/> <!--INT+FP+SFU First Lane Activation Power-->
+ <param name="static_cat4_addlane" value="0.6099397"/> <!--INT+FP+SFU Additional Lane Activation Power-->
+
+ <param name="static_cat5_flane" value="14.74826681"/> <!--INT+FP+TEX First Lane Activation Power-->
+ <param name="static_cat5_addlane" value="0.514367937"/> <!--INT+FP+TEX Additional Lane Activation Power-->
+
+ <param name="static_cat6_flane" value="48.94875596"/> <!--INT+FP+TENSOR First Lane Activation Power-->
+ <param name="static_cat6_addlane" value="0.0"/> <!--INT+FP+TENSOR Additional Lane Activation Power-->
+
+ <param name="static_light_flane" value="1.965373811"/> <!--LIGHT_SM First Lane Activation Power-->
+ <param name="static_light_addlane" value="0.003966868"/> <!--LIGHT_SM Additional Lane Activation Power-->
+
+ <param name="static_intadd_flane" value="19.70468506"/> <!--INT ADD First Lane Activation Power-->
+ <param name="static_intadd_addlane" value="0.388578623"/> <!--INT ADD Additional Lane Activation Power-->
+
+ <param name="static_intmul_flane" value="16.64811823"/> <!--INT MUL First Lane Activation Power-->
+ <param name="static_intmul_addlane" value="0.281803166"/> <!--INT MUL Additional Lane Activation Power-->
+
+ <param name="static_geomean_flane" value="17.21745077"/> <!--GEOMEAN First Lane Activation Power-->
+ <param name="static_geomean_addlane" value="0.650630555"/> <!--GEOMEAN Additional Lane Activation Power-->
+ <param name="static_shared_flane" value="31.40965691"/> <!--Shared Memory First Lane Activation Power-->
+ <param name="static_l1_flane" value="34.79491352"/> <!--L1D First Lane Activation Power-->
+ <param name="static_l2_flane" value="17.30654755"/> <!--L2D ADD+MUL First Lane Activation Power-->
+
+
+ <!-- LEAVE BELOW UNCHANGED! These are parameters used for initial GPUWattch model left untouched -->
+
+
+ <param name="GPU_Architecture" value="1"/><!-- 0-G80; 1-Fermi; others not supported -->
+ <param name="architecture" value="1"/> <!-- fermi:1 quadro:2 other: undefined-->
+ <param name="number_of_L1Directories" value="0"/>
+ <param name="number_of_L2Directories" value="0"/>
+ <param name="number_of_L2s" value="1"/> <!-- This number means how many L2 clusters in each cluster there can be multiple banks/ports -->
+ <param name="number_of_L3s" value="0"/> <!-- This number means how many L3 clusters -->
+ <param name="number_of_NoCs" value="1"/>
+ <param name="homogeneous_cores" value="1"/><!--1 means homo -->
+ <param name="homogeneous_L2s" value="1"/>
+ <param name="homogeneous_L1Directorys" value="1"/>
+ <param name="homogeneous_L2Directorys" value="1"/>
+ <param name="homogeneous_L3s" value="1"/>
+ <param name="homogeneous_ccs" value="1"/><!--cache coherece hardware -->
+ <param name="homogeneous_NoCs" value="1"/>
+ <param name="core_tech_node" value="23"/><!-- nm -->
+
+ <param name="number_cache_levels" value="2"/>
+ <param name="interconnect_projection_type" value="0"/><!--0: agressive wire technology; 1: conservative wire technology -->
+ <param name="device_type" value="0"/><!--0: HP(High Performance Type); 1: LSTP(Low standby power) 2: LOP (Low Operating Power) -->
+ <param name="longer_channel_device" value="1"/><!-- 0 no use; 1 use when possible -->
+ <param name="machine_bits" value="32"/>
+ <param name="virtual_address_width" value="32"/>
+ <param name="physical_address_width" value="32"/>
+ <param name="virtual_memory_page_size" value="4096"/>
+
+ <param name="IDLE_CORE_N" value="1"/>
+ <stat name="num_idle_cores" value="0"/><!-- Average Number of idle cores during this period -->
+ <stat name="total_cycles" value="total_cycles_match_mcpat"/>
+ <stat name="idle_cycles" value="idle_cycles_match_mcpat"/>
+ <stat name="busy_cycles" value="busy_cycles_match_mcpat"/>
+ <!--This page size(B) is complete different from the page size in Main memo secction. this page size is the size of
+ virtual memory from OS/Archi perspective; the page size in Main memo secction is the actuall physical line in a DRAM bank -->
+ <!-- *********************** cores ******************* -->
+ <component id="system.core0" name="core0">
+ <!-- Core property -->
+ <param name="clock_rate" value="700"/>
+ <param name="instruction_length" value="32"/>
+ <param name="opcode_width" value="9"/>
+ <!-- address width determins the tag_width in Cache, LSQ and buffers in cache controller
+ default value is machine_bits, if not set -->
+ <param name="machine_type" value="1"/><!-- 1 inorder; 0 OOO-->
+ <!-- inorder/OoO -->
+ <param name="number_hardware_threads" value="32"/>
+ <!-- number_instruction_fetch_ports(icache ports) is always 1 in single-thread processor,
+ it only may be more than one in SMT processors. BTB ports always equals to fetch ports since
+ branch information in consective branch instructions in the same fetch group can be read out from BTB once.-->
+ <param name="fetch_width" value="1"/>
+ <!-- fetch_width determins the size of cachelines of L1 cache block -->
+ <param name="number_instruction_fetch_ports" value="1"/>
+ <param name="decode_width" value="1"/>
+ <!-- decode_width determins the number of ports of the
+ renaming table (both RAM and CAM) scheme -->
+ <param name="issue_width" value="2"/>
+ <!-- issue_width determins the number of ports of Issue window and other logic
+ as in the complexity effective proccessors paper; issue_width==dispatch_width -->
+ <param name="commit_width" value="2"/>
+ <!-- commit_width determins the number of ports of register files -->
+ <param name="fp_issue_width" value="1"/>
+ <param name="prediction_width" value="0"/>
+ <!-- number of branch instructions can be predicted simultannouesl-->
+ <!-- Current version of McPAT does not distinguish int and floating point pipelines
+ Theses parameters are reserved for future use.-->
+ <param name="pipelines_per_core" value="1,1"/>
+ <!--integer_pipeline and floating_pipelines, if the floating_pipelines is 0, then the pipeline is shared-->
+ <param name="pipeline_depth" value="8,8"/>
+ <!-- pipeline depth of int and fp, if pipeline is shared, the second number is the average cycles of fp ops -->
+ <!-- issue and exe unit-->
+ <param name="ALU_per_core" value="32"/>
+ <!-- contains an adder, a shifter, and a logical unit -->
+ <param name="MUL_per_core" value="4"/>
+ <!-- For MUL and Div -->
+ <param name="FPU_per_core" value="32"/>
+ <!-- buffer between IF and ID stage -->
+ <param name="instruction_buffer_size" value="1"/>
+ <!-- buffer between ID and sche/exe stage -->
+ <param name="decoded_stream_buffer_size" value="1"/>
+ <param name="instruction_window_scheme" value="0"/><!-- 0 PHYREG based, 1 RSBASED-->
+ <!-- McPAT support 2 types of OoO cores, RS based and physical reg based-->
+ <param name="instruction_window_size" value="1"/>
+ <param name="fp_instruction_window_size" value="1"/>
+ <!-- the instruction issue Q as in Alpha 21264; The RS as in Intel P6 -->
+ <param name="ROB_size" value="0"/>
+ <!-- each in-flight instruction has an entry in ROB -->
+ <!-- registers -->
+ <!-- SM parameters Added by Syed Gilani -->
+ <param name="rf_banks" value="32"/>
+ <param name="simd_width" value="32"/>
+ <param name="collector_units" value="32"/>
+ <param name="core_clock_ratio" value="2"/>
+ <param name="warp_size" value="32"/>
+
+ <param name="archi_Regs_IRF_size" value="65536"/>
+ <param name="archi_Regs_FRF_size" value="32"/>
+ <!-- if OoO processor, phy_reg number is needed for renaming logic,
+ renaming logic is for both integer and floating point insts. -->
+ <param name="phy_Regs_IRF_size" value="32"/>
+ <param name="phy_Regs_FRF_size" value="32"/>
+ <!-- rename logic -->
+ <param name="rename_scheme" value="0"/>
+ <!-- can be RAM based(0) or CAM based(1) rename scheme
+ RAM-based scheme will have free list, status table;
+ CAM-based scheme have the valid bit in the data field of the CAM
+ both RAM and CAM need RAM-based checkpoint table, checkpoint_depth=# of in_flight instructions;
+ Detailed RAT Implementation see TR -->
+ <param name="register_windows_size" value="0"/>
+ <!-- how many windows in the windowed register file, sun processors;
+ no register windowing is used when this number is 0 -->
+ <!-- In OoO cores, loads and stores can be issued whether inorder(Pentium Pro) or (OoO)out-of-order(Alpha),
+ They will always try to exeute out-of-order though. -->
+ <param name="LSU_order" value="inorder"/>
+ <param name="store_buffer_size" value="32"/>
+ <!-- By default, in-order cores do not have load buffers -->
+ <param name="load_buffer_size" value="32"/>
+ <!-- number of ports refer to sustainable concurrent memory accesses -->
+ <param name="memory_ports" value="2"/>
+ <!-- max_allowed_in_flight_memo_instructions determins the # of ports of load and store buffer
+ as well as the ports of Dcache which is connected to LSU -->
+ <!-- dual-pumped Dcache can be used to save the extra read/write ports -->
+ <param name="RAS_size" value="1"/>
+ <!-- general stats, defines simulation periods;require total, idle, and busy cycles for senity check -->
+ <!-- please note: if target architecture is X86, then all the instrucions refer to (fused) micro-ops -->
+ <stat name="total_instructions" value="total_instructions_match_mcpat"/>
+ <stat name="int_instructions" value="int_instruction_match_mcpat"/>
+ <stat name="fp_instructions" value="flt_instruction_match_mcpat"/>
+ <stat name="branch_instructions" value="branch_instruction_match_mcpat"/>
+ <stat name="branch_mispredictions" value="0"/>
+ <stat name="load_instructions" value="load_instruction_match_mcpat"/>
+ <stat name="store_instructions" value="store_instruction_match_mcpat"/>
+ <stat name="committed_instructions" value="total_instructions_match_mcpat"/>
+ <stat name="committed_int_instructions" value="int_instruction_match_mcpat"/>
+ <stat name="committed_fp_instructions" value="flt_instruction_match_mcpat"/>
+ <stat name="pipeline_duty_cycle" value="0.6"/><!--<=1, runtime_ipc/peak_ipc; averaged for all cores if homogenous -->
+ <!-- the following cycle stats are used for heterogeneouse cores only,
+ please ignore them if homogeneouse cores -->
+ <stat name="total_cycles" value="total_cycles_match_mcpat"/>
+ <stat name="idle_cycles" value="idle_cycles_match_mcpat"/>
+ <stat name="busy_cycles" value="busy_cycles_match_mcpat"/>
+ <!-- instruction buffer stats -->
+ <!-- ROB stats, both RS and Phy based OoOs have ROB
+ performance simulator should capture the difference on accesses,
+ otherwise, McPAT has to guess based on number of commited instructions. -->
+ <stat name="ROB_reads" value="263886"/>
+ <stat name="ROB_writes" value="263886"/>
+ <!-- RAT accesses -->
+ <stat name="rename_accesses" value="263886"/>
+ <stat name="fp_rename_accesses" value="263886"/>
+ <!-- decode and rename stage use this, should be total ic - nop -->
+ <!-- Inst window stats -->
+ <stat name="inst_window_reads" value="263886"/>
+ <stat name="inst_window_writes" value="263886"/>
+ <stat name="inst_window_wakeup_accesses" value="263886"/>
+ <stat name="fp_inst_window_reads" value="263886"/>
+ <stat name="fp_inst_window_writes" value="263886"/>
+ <stat name="fp_inst_window_wakeup_accesses" value="263886"/>
+ <!-- RF accesses -->
+ <stat name="int_regfile_reads" value="int_register_read_access_match_mcpat"/>
+ <stat name="float_regfile_reads" value="int_register_write_access_match_mcpat"/>
+ <stat name="int_regfile_writes" value="float_register_read_access_match_mcpat"/>
+ <stat name="float_regfile_writes" value="float_register_write_access_match_mcpat"/>
+
+ <!-- The following stat is for operand collector power - Added by Syed -->
+ <stat name="non_rf_operands" value="0"/>
+
+ <!-- accesses to the working reg -->
+ <stat name="function_calls" value="0"/>
+ <stat name="context_switches" value="0"/> <!--not used in the McPAT -->
+ <!-- Number of Windowes switches (number of function calls and returns)-->
+ <!-- Alu stats by default, the processor has one FPU that includes the divider and
+ multiplier. The fpu accesses should include accesses to multiplier and divider -->
+ <stat name="ialu_accesses" value="ialu_accesses_match_mcpat"/>
+ <stat name="fpu_accesses" value="fpu_accesses_match_mcpat"/>
+ <stat name="mul_accesses" value="mul_accesses_match_mcpat"/>
+ <stat name="cdb_alu_accesses" value="0"/>
+ <stat name="cdb_mul_accesses" value="0"/>
+ <stat name="cdb_fpu_accesses" value="0"/>
+ <!-- multiple cycle accesses should be counted multiple times,
+ otherwise, McPAT can use internal counter for different floating point instructions
+ to get final accesses. But that needs detailed info for floating point inst mix -->
+ <!-- currently the performance simulator should
+ make sure all the numbers are final numbers,
+ including the explicit read/write accesses,
+ and the implicite accesses such as replacements and etc.
+ Future versions of McPAT may be able to reason the implicite access
+ based on param and stats of last level cache
+ The same rule applies to all cache access stats too! -->
+ <!-- following is AF for max power computation.
+ Do not change them, unless you understand them-->
+ <stat name="IFU_duty_cycle" value="0.25"/>
+ <stat name="LSU_duty_cycle" value="0.25"/>
+ <stat name="MemManU_I_duty_cycle" value="1"/>
+ <stat name="MemManU_D_duty_cycle" value="0.25"/>
+ <stat name="ALU_duty_cycle" value="0.9"/>
+ <stat name="MUL_duty_cycle" value="0.5"/>
+ <stat name="FPU_duty_cycle" value="1"/><!-- FPU numbers are already average -->
+ <stat name="ALU_cdb_duty_cycle" value="0.9"/>
+ <stat name="MUL_cdb_duty_cycle" value="0.5"/>
+ <stat name="FPU_cdb_duty_cycle" value="15"/>
+ <component id="system.core0.predictor" name="PBT">
+ <!-- branch predictor; tournament predictor see Alpha implementation -->
+ <param name="local_predictor_size" value="10,3"/>
+ <param name="local_predictor_entries" value="1024"/>
+ <param name="global_predictor_entries" value="4096"/>
+ <param name="global_predictor_bits" value="2"/>
+ <param name="chooser_predictor_entries" value="4096"/>
+ <param name="chooser_predictor_bits" value="2"/>
+ <!-- These parameters can be combined like below in next version
+ <param name="load_predictor" value="10,3,1024"/>
+ <param name="global_predictor" value="4096,2"/>
+ <param name="predictor_chooser" value="4096,2"/>
+ -->
+ </component>
+ <component id="system.core0.itlb" name="itlb">
+ <param name="number_entries" value="1"/>
+ <stat name="total_accesses" value="0"/>
+ <stat name="total_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <!-- there is no write requests to itlb although writes happen to itlb after miss,
+ which is actually a replacement -->
+ </component>
+ <component id="system.core0.icache" name="icache">
+ <!-- there is no write requests to itlb although writes happen to it after miss,
+ which is actually a replacement -->
+ <param name="icache_config" value="16384,128,4,1,1,3,8,0"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <!-- cache_policy;//0 no write or write-though with non-write allocate;1 write-back with write-allocate -->
+ <param name="buffer_sizes" value="16, 16, 16,0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="total_instructions_match_mcpat"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.dtlb" name="dtlb">
+ <param name="number_entries" value="1"/>
+ <stat name="total_accesses" value="0"/>
+ <stat name="total_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.ccache" name="ccache">
+ <!-- all the buffer related are optional -->
+ <param name="ccache_config" value="16384,64,2,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="ccache_read_accesses_match_mcpat"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="ccache_read_misses_match_mcpat"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.tcache" name="tcache">
+ <!-- all the buffer related are optional -->
+ <param name="tcache_config" value="49152,128,8,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="tcache_read_accesses_match_mcpat"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="tcache_read_misses_match_mcpat"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <!--model the shared memory by mimicing dcache-->
+ <component id="system.core0.sharedmemory" name="sharedmemory">
+ <!-- all the buffer related are optional -->
+ <param name="sharedmemory_config" value="98304,16,1,16,1,3,16,0"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="sharedmemory_read_access_match_mcpat"/>
+ <stat name="write_accesses" value="sharedmemory_write_access_match_mcpat"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.dcache" name="dcache">
+ <!-- all the buffer related are optional -->
+ <param name="dcache_config" value="16384,32,4,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="dcache_read_access_match_mcpat"/>
+ <stat name="write_accesses" value="dcache_write_access_match_mcpat"/>
+ <stat name="read_misses" value="dcache_read_miss_match_mcpat"/>
+ <stat name="write_misses" value="dcache_write_miss_match_mcpat"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.BTB" name="BTB">
+ <!-- all the buffer related are optional -->
+ <param name="BTB_config" value="8192,4,2,1, 1,3"/>
+ <!-- the parameters are capacity,block_width,associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ </component>
+ </component>
+ <component id="system.L1Directory0" name="L1Directory0">
+ <param name="Directory_type" value="0"/>
+ <!--0 cam based shadowed tag. 1 directory cache -->
+ <param name="Dir_config" value="2048,1,0,1, 4, 4,8"/>
+ <!-- the parameters are capacity,block_width, associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ <param name="buffer_sizes" value="8, 8, 8, 8"/>
+ <!-- all the buffer related are optional -->
+ <param name="clockrate" value="1400"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw search ports -->
+ <param name="device_type" value="0"/>
+ <!-- altough there are multiple access types,
+ Performance simulator needs to cast them into reads or writes
+ e.g. the invalidates can be considered as writes -->
+ <stat name="read_accesses" value="800000"/>
+ <stat name="write_accesses" value="27276"/>
+ <stat name="read_misses" value="1632"/>
+ <stat name="write_misses" value="183"/>
+ <stat name="conflicts" value="20"/>
+ <stat name="duty_cycle" value="0.45"/>
+ </component>
+ <component id="system.L2Directory0" name="L2Directory0">
+ <param name="Directory_type" value="1"/>
+ <!--0 cam based shadowed tag. 1 directory cache -->
+ <param name="Dir_config" value="1048576,16,16,1,2, 100"/>
+ <!-- the parameters are capacity,block_width, associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ <param name="buffer_sizes" value="8, 8, 8, 8"/>
+ <!-- all the buffer related are optional -->
+ <param name="clockrate" value="1400"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw search ports -->
+ <param name="device_type" value="0"/>
+ <!-- altough there are multiple access types,
+ Performance simulator needs to cast them into reads or writes
+ e.g. the invalidates can be considered as writes -->
+ <stat name="read_accesses" value="0"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.45"/>
+ </component>
+ <component id="system.L20" name="L20">
+ <!-- all the buffer related are optional -->
+ <param name="L2_config" value="131072,128,16,1, 4,23, 64, 1"/>
+ <!-- consider 4-way bank interleaving for Niagara 1 -->
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <param name="clockrate" value="2962"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw ports -->
+ <param name="device_type" value="0"/>
+ <stat name="read_accesses" value="200000"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.5"/>
+ </component>
+
+<!--**********************************************************************-->
+<component id="system.L30" name="L30">
+ <param name="L3_config" value="1048576,64,16,1, 2,100, 64,1"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="clockrate" value="3500"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw ports -->
+ <param name="device_type" value="0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="58824"/>
+ <stat name="write_accesses" value="27276"/>
+ <stat name="read_misses" value="1632"/>
+ <stat name="write_misses" value="183"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.35"/>
+ </component>
+
+
+<!--**********************************************************************-->
+ <component id="system.NoC0" name="noc0">
+ <param name="clockrate" value="700"/>
+ <param name="type" value="1"/>
+ <!-- 1 NoC, O bus -->
+ <param name="horizontal_nodes" value="2"/>
+ <param name="vertical_nodes" value="1"/>
+ <param name="has_global_link" value="0"/>
+ <!-- 1 has global link, 0 does not have global link -->
+ <param name="link_throughput" value="1"/><!--w.r.t clock -->
+ <param name="link_latency" value="1"/><!--w.r.t clock -->
+ <!-- througput >= latency -->
+ <!-- Router architecture -->
+ <param name="input_ports" value="6"/>
+ <param name="output_ports" value="6"/>
+ <param name="virtual_channel_per_port" value="1"/>
+ <!-- input buffer; in classic routers only input ports need buffers -->
+ <param name="flit_bits" value="32"/>
+ <param name="input_buffer_entries_per_vc" value="1"/><!--VCs within the same ports share input buffers whose size is propotional to the number of VCs-->
+ <param name="chip_coverage" value="1"/>
+ <!-- When multiple NOC present, one NOC will cover part of the whole chip. chip_coverage <=1 -->
+ <stat name="total_accesses" value="0"/>
+ <!-- This is the number of total accesses within the whole network not for each router -->
+ <stat name="duty_cycle" value="0.6"/>
+ </component>
+<!--**********************************************************************-->
+<!--**********************************************************************-->
+
+ <component id="system.mem" name="mem">
+ <!-- Main memory property -->
+ <param name="mem_tech_node" value="23"/>
+ <param name="device_clock" value="200"/><!--MHz, this is clock rate of the actual memory device, not the FSB -->
+ <param name="peak_transfer_rate" value="3200"/><!--MB/S-->
+ <param name="internal_prefetch_of_DRAM_chip" value="4"/>
+ <!-- 2 for DDR, 4 for DDR2, 8 for DDR3...-->
+ <!-- the device clock, peak_transfer_rate, and the internal prefetch decide the DIMM property -->
+ <!-- above numbers can be easily found from Wikipedia -->
+ <param name="capacity_per_channel" value="4096"/> <!-- MB -->
+ <!-- capacity_per_Dram_chip=capacity_per_channel/number_of_dimms/number_ranks/Dram_chips_per_rank
+ Current McPAT assumes single DIMMs are used.-->
+ <param name="number_ranks" value="2"/>
+ <param name="num_banks_of_DRAM_chip" value="6"/>
+ <param name="Block_width_of_DRAM_chip" value="64"/> <!-- B -->
+ <param name="output_width_of_DRAM_chip" value="8"/>
+ <!--number of Dram_chips_per_rank=" 72/output_width_of_DRAM_chip-->
+ <!--number of Dram_chips_per_rank=" 72/output_width_of_DRAM_chip-->
+ <param name="page_size_of_DRAM_chip" value="8"/> <!-- 8 or 16 -->
+ <param name="burstlength_of_DRAM_chip" value="8"/>
+ <stat name="memory_accesses" value="1052"/>
+ <stat name="memory_reads" value="1052"/>
+ <stat name="memory_writes" value="1052"/>
+ </component>
+ <component id="system.mc" name="mc">
+ <!-- Memeory controllers are for DDR(2,3...) DIMMs -->
+ <!-- current version of McPAT uses published values for base parameters of memory controller
+ improvments on MC will be added in later versions. -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="mc_clock" value="1848"/><!--DIMM IO bus clock rate MHz DDR2-400 for Niagara 1-->
+ <param name="peak_transfer_rate" value="29568"/><!--MB/S Syed: GTX 470 has 177.4GB/s mem transfer rate with 6 MCs -->
+ <param name="block_size" value="64"/><!--B-->
+ <param name="number_mcs" value="6"/><!-- 6 GDDR5 memory controllers -->
+ <!-- current McPAT only supports homogeneous memory controllers -->
+ <param name="memory_channels_per_mc" value="2"/>
+ <param name="number_ranks" value="1"/>
+ <param name="withPHY" value="0"/>
+ <!-- # of ranks of each channel-->
+ <param name="req_window_size_per_channel" value="16"/>
+ <param name="IO_buffer_size_per_channel" value="16"/>
+ <param name="databus_width" value="32"/>
+ <param name="addressbus_width" value="32"/>
+ <param name="PRT_entries" value="32"/>
+ <!-- # of empirical DRAM model parameter -->
+ <param name="dram_cmd_coeff" value="0"/>
+ <param name="dram_act_coeff" value="0"/>
+ <param name="dram_nop_coeff" value="0"/>
+ <param name="dram_activity_coeff" value="0"/>
+ <param name="dram_pre_coeff" value="3.8475e-8f"/>
+ <param name="dram_rd_coeff" value="7.74707143e-8f"/>
+ <param name="dram_wr_coeff" value="3.54664286e-8f"/>
+ <param name="dram_req_coeff" value="0"/>
+ <param name="dram_const_coeff" value="0"/>
+
+ <!-- McPAT will add the control bus width to the addressbus width automatically -->
+ <stat name="memory_accesses" value="memory_accesses_match_mcpat"/>
+ <stat name="memory_reads" value="memory_reads_match_mcpat"/>
+ <stat name="memory_writes" value="memory_writes_match_mcpat"/>
+ <!-- McPAT does not track individual mc, instead, it takes the total accesses and calculate
+ the average power per MC or per channel. This is sufficent for most application.
+ Further trackdown can be easily added in later versions. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.niu" name="niu">
+ <!-- On chip 10Gb Ethernet NIC, including XAUI Phy and MAC controller -->
+ <!-- For a minimum IP packet size of 84B at 10Gb/s, a new packet arrives every 67.2ns.
+ the low bound of clock rate of a 10Gb MAC is 150Mhz -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="clockrate" value="350"/>
+ <param name="number_units" value="0"/> <!-- unlike PCIe and memory controllers, each Ethernet controller only have one port -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- ratio of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual nic, instead, it takes the total accesses and calculate
+ the average power per nic or per channel. This is sufficent for most application. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.pcie" name="pcie">
+ <!-- On chip PCIe controller, including Phy-->
+ <!-- For a minimum PCIe packet size of 84B at 8Gb/s per lane (PCIe 3.0), a new packet arrives every 84ns.
+ the low bound of clock rate of a PCIe per lane logic is 120Mhz -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="withPHY" value="1"/>
+ <param name="clockrate" value="350"/>
+ <param name="number_units" value="0"/>
+ <param name="num_channels" value="8"/> <!-- 2 ,4 ,8 ,16 ,32 -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- Percentage of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual pcie controllers, instead, it takes the total accesses and calculate
+ the average power per pcie controller or per channel. This is sufficent for most application. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.flashc" name="flashc">
+ <param name="number_flashcs" value="0"/>
+ <param name="type" value="1"/> <!-- 1: low power; 0 high performance -->
+ <param name="withPHY" value="1"/>
+ <param name="peak_transfer_rate" value="200"/><!--Per controller sustainable reak rate MB/S -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- Percentage of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual flash controller, instead, it takes the total accesses and calculate
+ the average power per fc or per channel. This is sufficent for most application -->
+ </component>
+<!--**********************************************************************-->
+
+ </component>
+</component>
\ No newline at end of file diff --git a/configs/tested-cfgs/SM7_GV100/accelwattch_sass_hybrid.xml b/configs/tested-cfgs/SM7_GV100/accelwattch_sass_hybrid.xml new file mode 100644 index 0000000..175f1fd --- /dev/null +++ b/configs/tested-cfgs/SM7_GV100/accelwattch_sass_hybrid.xml @@ -0,0 +1,613 @@ +<?xml version="1.0" ?>
+<component id="root" name="root">
+ <component id="system" name="system">
+<!-- Copyright (c) 2018-2021, Vijay Kandiah, Junrui Pan, Mahmoud Khairy, Scott Peverelle, Timothy Rogers, Tor M. Aamodt, Nikos Hardavellas
+Northwestern University, Purdue University, The University of British Columbia
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+ list of conditions and the following disclaimer;
+2. Redistributions in binary form must reproduce the above copyright notice,
+ this list of conditions and the following disclaimer in the documentation
+ and/or other materials provided with the distribution;
+3. Neither the names of Northwestern University, Purdue University,
+ The University of British Columbia nor the names of their contributors
+ may be used to endorse or promote products derived from this software
+ without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE. -->
+
+ <!-- DYNAMIC Power Activity Factors -->
+
+ <param name="TOT_INST" value="100" /> <!--Instruction Buffer Activity Factor(Total Warp Instructions)-->
+ <param name="FP_INT" value="48.259994" /> <!--Scheduler Activity Factor(Non-Memory Warp Instructons)-->
+ <param name="IC_H" value="0" /> <!--Instruction Cache Hits -->
+ <param name="IC_M" value="0" /> <!--Instruction Cache Misses-->
+ <param name="DC_RH" value="6.607568654" /> <!--L1 Data Cache Read Hits-->
+ <param name="DC_RM" value="7.359649633" /> <!--L1 Data Cache Read Misses-->
+ <param name="DC_WH" value="0.456620598" /> <!--L1 Data Cache Write Hits-->
+ <param name="DC_WM" value="11.87512" /> <!--L1 Data Cache Write Misses-->
+ <param name="CC_H" value="5.932232271" /> <!--Constant Cache Hits-->
+ <param name="CC_M" value="6.607445701" /> <!--Constant Cache Misses-->
+ <param name="SHRD_ACC" value="5.116439734" /> <!--Shared Memory Accesses-->
+ <param name="REG_RD" value="0" /> <!--Register Reads-->
+ <param name="REG_WR" value="0" /> <!--Register Writes-->
+ <param name="INT_ACC" value="7.725550311" /> <!--Integer ALU Accesses-->
+ <param name="FP_ACC" value="0.201142468" /> <!--FPU Accesses-->
+ <param name="DP_ACC" value="0.201146878" /> <!--Integer Accesses-->
+ <param name="INT_MUL_ACC" value="0.130624774" /> <!--Integer MUL Accesses-->
+ <param name="FP_MUL_ACC" value="0.130623287" /> <!--FP MUL Accesses-->
+ <param name="FP_SQRT_ACC" value="0.195089274" /> <!--SFU: sqrt Accesses-->
+ <param name="FP_LG_ACC" value="0.130617842" /> <!--SFU: log Accesses-->
+ <param name="FP_SIN_ACC" value="0.202205086" /> <!--SFU: sin/cos Accesses-->
+ <param name="FP_EXP_ACC" value="0.130621909" /> <!--SFU: exponent Accesses-->
+ <param name="DP_MUL_ACC" value="0.130622532" /> <!--DP MUL Accesses-->
+ <param name="TENSOR_ACC" value="0.240811904" /> <!--Tensor Accesses-->
+ <param name="TEX_ACC" value="3.453002628" /> <!--Texture Unit Accesses-->
+ <param name="MEM_RD" value="0.01527021" /> <!--DRAM Reads-->
+ <param name="MEM_WR" value="0.018509345" /> <!--DRAM Writes-->
+ <param name="MEM_PRE" value="0.00509007" /> <!--DRAM Precharge-->
+ <param name="L2_RH" value="3.391355385" /> <!--L2 Data Cache Read Hits-->
+ <param name="L2_RM" value="6.440581599" /> <!--L2 Data Cache Read Misses-->
+ <param name="L2_WH" value="11.09478657" /> <!--L2 Data Cache Write Hits-->
+ <param name="L2_WM" value="3.288716634" /> <!--L2 Data Cache Write Misses-->
+ <param name="NOC_A" value="86.3134871" /> <!--Interconnect Network Accesses-->
+ <param name="PIPE_A" value="0.514" /> <!--Pipeline Accesses-->
+
+ <!-- STATIC & CONSTANT Power -->
+
+ <param name="constant_power" value="32.32522272" /> <!--Constant power-->
+ <param name="idle_core_power" value="0.28279166"/> <!--Idle SM power-->
+
+ <param name="static_cat1_flane" value="15.29035866"/> <!--INT (ADD+MUL) First Lane Activation Power-->
+ <param name="static_cat1_addlane" value="0.586233603"/> <!--INT (ADD+MUL) Additional Lane Activation Power-->
+
+ <param name="static_cat2_flane" value="18.6179906"/> <!--INT+FP First Lane Activation Power-->
+ <param name="static_cat2_addlane" value="0.645228013"/> <!--INT+FP Additional Lane Activation Power-->
+
+ <param name="static_cat3_flane" value="19.10017723"/> <!--INT+FP+DP First Lane Activation Power-->
+ <param name="static_cat3_addlane" value="0.726863055"/> <!--INT+FP+DP Additional Lane Activation Power-->
+
+ <param name="static_cat4_flane" value="18.55029744"/> <!--INT+FP+SFU First Lane Activation Power-->
+ <param name="static_cat4_addlane" value="0.6099397"/> <!--INT+FP+SFU Additional Lane Activation Power-->
+
+ <param name="static_cat5_flane" value="14.74826681"/> <!--INT+FP+TEX First Lane Activation Power-->
+ <param name="static_cat5_addlane" value="0.514367937"/> <!--INT+FP+TEX Additional Lane Activation Power-->
+
+ <param name="static_cat6_flane" value="48.94875596"/> <!--INT+FP+TENSOR First Lane Activation Power-->
+ <param name="static_cat6_addlane" value="0.0"/> <!--INT+FP+TENSOR Additional Lane Activation Power-->
+
+ <param name="static_light_flane" value="1.965373811"/> <!--LIGHT_SM First Lane Activation Power-->
+ <param name="static_light_addlane" value="0.003966868"/> <!--LIGHT_SM Additional Lane Activation Power-->
+
+ <param name="static_intadd_flane" value="19.70468506"/> <!--INT ADD First Lane Activation Power-->
+ <param name="static_intadd_addlane" value="0.388578623"/> <!--INT ADD Additional Lane Activation Power-->
+
+ <param name="static_intmul_flane" value="16.64811823"/> <!--INT MUL First Lane Activation Power-->
+ <param name="static_intmul_addlane" value="0.281803166"/> <!--INT MUL Additional Lane Activation Power-->
+
+ <param name="static_geomean_flane" value="17.21745077"/> <!--GEOMEAN First Lane Activation Power-->
+ <param name="static_geomean_addlane" value="0.650630555"/> <!--GEOMEAN Additional Lane Activation Power-->
+ <param name="static_shared_flane" value="31.40965691"/> <!--Shared Memory First Lane Activation Power-->
+ <param name="static_l1_flane" value="34.79491352"/> <!--L1D First Lane Activation Power-->
+ <param name="static_l2_flane" value="17.30654755"/> <!--L2D ADD+MUL First Lane Activation Power-->
+
+
+ <!-- LEAVE BELOW UNCHANGED! These are parameters used for initial GPUWattch model left untouched -->
+
+
+ <param name="GPU_Architecture" value="1"/><!-- 0-G80; 1-Fermi; others not supported -->
+ <param name="architecture" value="1"/> <!-- fermi:1 quadro:2 other: undefined-->
+ <param name="number_of_L1Directories" value="0"/>
+ <param name="number_of_L2Directories" value="0"/>
+ <param name="number_of_L2s" value="1"/> <!-- This number means how many L2 clusters in each cluster there can be multiple banks/ports -->
+ <param name="number_of_L3s" value="0"/> <!-- This number means how many L3 clusters -->
+ <param name="number_of_NoCs" value="1"/>
+ <param name="homogeneous_cores" value="1"/><!--1 means homo -->
+ <param name="homogeneous_L2s" value="1"/>
+ <param name="homogeneous_L1Directorys" value="1"/>
+ <param name="homogeneous_L2Directorys" value="1"/>
+ <param name="homogeneous_L3s" value="1"/>
+ <param name="homogeneous_ccs" value="1"/><!--cache coherece hardware -->
+ <param name="homogeneous_NoCs" value="1"/>
+ <param name="core_tech_node" value="23"/><!-- nm -->
+
+ <param name="number_cache_levels" value="2"/>
+ <param name="interconnect_projection_type" value="0"/><!--0: agressive wire technology; 1: conservative wire technology -->
+ <param name="device_type" value="0"/><!--0: HP(High Performance Type); 1: LSTP(Low standby power) 2: LOP (Low Operating Power) -->
+ <param name="longer_channel_device" value="1"/><!-- 0 no use; 1 use when possible -->
+ <param name="machine_bits" value="32"/>
+ <param name="virtual_address_width" value="32"/>
+ <param name="physical_address_width" value="32"/>
+ <param name="virtual_memory_page_size" value="4096"/>
+
+ <param name="IDLE_CORE_N" value="1"/>
+ <stat name="num_idle_cores" value="0"/><!-- Average Number of idle cores during this period -->
+ <stat name="total_cycles" value="total_cycles_match_mcpat"/>
+ <stat name="idle_cycles" value="idle_cycles_match_mcpat"/>
+ <stat name="busy_cycles" value="busy_cycles_match_mcpat"/>
+ <!--This page size(B) is complete different from the page size in Main memo secction. this page size is the size of
+ virtual memory from OS/Archi perspective; the page size in Main memo secction is the actuall physical line in a DRAM bank -->
+ <!-- *********************** cores ******************* -->
+ <component id="system.core0" name="core0">
+ <!-- Core property -->
+ <param name="clock_rate" value="700"/>
+ <param name="instruction_length" value="32"/>
+ <param name="opcode_width" value="9"/>
+ <!-- address width determins the tag_width in Cache, LSQ and buffers in cache controller
+ default value is machine_bits, if not set -->
+ <param name="machine_type" value="1"/><!-- 1 inorder; 0 OOO-->
+ <!-- inorder/OoO -->
+ <param name="number_hardware_threads" value="32"/>
+ <!-- number_instruction_fetch_ports(icache ports) is always 1 in single-thread processor,
+ it only may be more than one in SMT processors. BTB ports always equals to fetch ports since
+ branch information in consective branch instructions in the same fetch group can be read out from BTB once.-->
+ <param name="fetch_width" value="1"/>
+ <!-- fetch_width determins the size of cachelines of L1 cache block -->
+ <param name="number_instruction_fetch_ports" value="1"/>
+ <param name="decode_width" value="1"/>
+ <!-- decode_width determins the number of ports of the
+ renaming table (both RAM and CAM) scheme -->
+ <param name="issue_width" value="2"/>
+ <!-- issue_width determins the number of ports of Issue window and other logic
+ as in the complexity effective proccessors paper; issue_width==dispatch_width -->
+ <param name="commit_width" value="2"/>
+ <!-- commit_width determins the number of ports of register files -->
+ <param name="fp_issue_width" value="1"/>
+ <param name="prediction_width" value="0"/>
+ <!-- number of branch instructions can be predicted simultannouesl-->
+ <!-- Current version of McPAT does not distinguish int and floating point pipelines
+ Theses parameters are reserved for future use.-->
+ <param name="pipelines_per_core" value="1,1"/>
+ <!--integer_pipeline and floating_pipelines, if the floating_pipelines is 0, then the pipeline is shared-->
+ <param name="pipeline_depth" value="8,8"/>
+ <!-- pipeline depth of int and fp, if pipeline is shared, the second number is the average cycles of fp ops -->
+ <!-- issue and exe unit-->
+ <param name="ALU_per_core" value="32"/>
+ <!-- contains an adder, a shifter, and a logical unit -->
+ <param name="MUL_per_core" value="4"/>
+ <!-- For MUL and Div -->
+ <param name="FPU_per_core" value="32"/>
+ <!-- buffer between IF and ID stage -->
+ <param name="instruction_buffer_size" value="1"/>
+ <!-- buffer between ID and sche/exe stage -->
+ <param name="decoded_stream_buffer_size" value="1"/>
+ <param name="instruction_window_scheme" value="0"/><!-- 0 PHYREG based, 1 RSBASED-->
+ <!-- McPAT support 2 types of OoO cores, RS based and physical reg based-->
+ <param name="instruction_window_size" value="1"/>
+ <param name="fp_instruction_window_size" value="1"/>
+ <!-- the instruction issue Q as in Alpha 21264; The RS as in Intel P6 -->
+ <param name="ROB_size" value="0"/>
+ <!-- each in-flight instruction has an entry in ROB -->
+ <!-- registers -->
+ <!-- SM parameters Added by Syed Gilani -->
+ <param name="rf_banks" value="32"/>
+ <param name="simd_width" value="32"/>
+ <param name="collector_units" value="32"/>
+ <param name="core_clock_ratio" value="2"/>
+ <param name="warp_size" value="32"/>
+
+ <param name="archi_Regs_IRF_size" value="65536"/>
+ <param name="archi_Regs_FRF_size" value="32"/>
+ <!-- if OoO processor, phy_reg number is needed for renaming logic,
+ renaming logic is for both integer and floating point insts. -->
+ <param name="phy_Regs_IRF_size" value="32"/>
+ <param name="phy_Regs_FRF_size" value="32"/>
+ <!-- rename logic -->
+ <param name="rename_scheme" value="0"/>
+ <!-- can be RAM based(0) or CAM based(1) rename scheme
+ RAM-based scheme will have free list, status table;
+ CAM-based scheme have the valid bit in the data field of the CAM
+ both RAM and CAM need RAM-based checkpoint table, checkpoint_depth=# of in_flight instructions;
+ Detailed RAT Implementation see TR -->
+ <param name="register_windows_size" value="0"/>
+ <!-- how many windows in the windowed register file, sun processors;
+ no register windowing is used when this number is 0 -->
+ <!-- In OoO cores, loads and stores can be issued whether inorder(Pentium Pro) or (OoO)out-of-order(Alpha),
+ They will always try to exeute out-of-order though. -->
+ <param name="LSU_order" value="inorder"/>
+ <param name="store_buffer_size" value="32"/>
+ <!-- By default, in-order cores do not have load buffers -->
+ <param name="load_buffer_size" value="32"/>
+ <!-- number of ports refer to sustainable concurrent memory accesses -->
+ <param name="memory_ports" value="2"/>
+ <!-- max_allowed_in_flight_memo_instructions determins the # of ports of load and store buffer
+ as well as the ports of Dcache which is connected to LSU -->
+ <!-- dual-pumped Dcache can be used to save the extra read/write ports -->
+ <param name="RAS_size" value="1"/>
+ <!-- general stats, defines simulation periods;require total, idle, and busy cycles for senity check -->
+ <!-- please note: if target architecture is X86, then all the instrucions refer to (fused) micro-ops -->
+ <stat name="total_instructions" value="total_instructions_match_mcpat"/>
+ <stat name="int_instructions" value="int_instruction_match_mcpat"/>
+ <stat name="fp_instructions" value="flt_instruction_match_mcpat"/>
+ <stat name="branch_instructions" value="branch_instruction_match_mcpat"/>
+ <stat name="branch_mispredictions" value="0"/>
+ <stat name="load_instructions" value="load_instruction_match_mcpat"/>
+ <stat name="store_instructions" value="store_instruction_match_mcpat"/>
+ <stat name="committed_instructions" value="total_instructions_match_mcpat"/>
+ <stat name="committed_int_instructions" value="int_instruction_match_mcpat"/>
+ <stat name="committed_fp_instructions" value="flt_instruction_match_mcpat"/>
+ <stat name="pipeline_duty_cycle" value="0.6"/><!--<=1, runtime_ipc/peak_ipc; averaged for all cores if homogenous -->
+ <!-- the following cycle stats are used for heterogeneouse cores only,
+ please ignore them if homogeneouse cores -->
+ <stat name="total_cycles" value="total_cycles_match_mcpat"/>
+ <stat name="idle_cycles" value="idle_cycles_match_mcpat"/>
+ <stat name="busy_cycles" value="busy_cycles_match_mcpat"/>
+ <!-- instruction buffer stats -->
+ <!-- ROB stats, both RS and Phy based OoOs have ROB
+ performance simulator should capture the difference on accesses,
+ otherwise, McPAT has to guess based on number of commited instructions. -->
+ <stat name="ROB_reads" value="263886"/>
+ <stat name="ROB_writes" value="263886"/>
+ <!-- RAT accesses -->
+ <stat name="rename_accesses" value="263886"/>
+ <stat name="fp_rename_accesses" value="263886"/>
+ <!-- decode and rename stage use this, should be total ic - nop -->
+ <!-- Inst window stats -->
+ <stat name="inst_window_reads" value="263886"/>
+ <stat name="inst_window_writes" value="263886"/>
+ <stat name="inst_window_wakeup_accesses" value="263886"/>
+ <stat name="fp_inst_window_reads" value="263886"/>
+ <stat name="fp_inst_window_writes" value="263886"/>
+ <stat name="fp_inst_window_wakeup_accesses" value="263886"/>
+ <!-- RF accesses -->
+ <stat name="int_regfile_reads" value="int_register_read_access_match_mcpat"/>
+ <stat name="float_regfile_reads" value="int_register_write_access_match_mcpat"/>
+ <stat name="int_regfile_writes" value="float_register_read_access_match_mcpat"/>
+ <stat name="float_regfile_writes" value="float_register_write_access_match_mcpat"/>
+
+ <!-- The following stat is for operand collector power - Added by Syed -->
+ <stat name="non_rf_operands" value="0"/>
+
+ <!-- accesses to the working reg -->
+ <stat name="function_calls" value="0"/>
+ <stat name="context_switches" value="0"/> <!--not used in the McPAT -->
+ <!-- Number of Windowes switches (number of function calls and returns)-->
+ <!-- Alu stats by default, the processor has one FPU that includes the divider and
+ multiplier. The fpu accesses should include accesses to multiplier and divider -->
+ <stat name="ialu_accesses" value="ialu_accesses_match_mcpat"/>
+ <stat name="fpu_accesses" value="fpu_accesses_match_mcpat"/>
+ <stat name="mul_accesses" value="mul_accesses_match_mcpat"/>
+ <stat name="cdb_alu_accesses" value="0"/>
+ <stat name="cdb_mul_accesses" value="0"/>
+ <stat name="cdb_fpu_accesses" value="0"/>
+ <!-- multiple cycle accesses should be counted multiple times,
+ otherwise, McPAT can use internal counter for different floating point instructions
+ to get final accesses. But that needs detailed info for floating point inst mix -->
+ <!-- currently the performance simulator should
+ make sure all the numbers are final numbers,
+ including the explicit read/write accesses,
+ and the implicite accesses such as replacements and etc.
+ Future versions of McPAT may be able to reason the implicite access
+ based on param and stats of last level cache
+ The same rule applies to all cache access stats too! -->
+ <!-- following is AF for max power computation.
+ Do not change them, unless you understand them-->
+ <stat name="IFU_duty_cycle" value="0.25"/>
+ <stat name="LSU_duty_cycle" value="0.25"/>
+ <stat name="MemManU_I_duty_cycle" value="1"/>
+ <stat name="MemManU_D_duty_cycle" value="0.25"/>
+ <stat name="ALU_duty_cycle" value="0.9"/>
+ <stat name="MUL_duty_cycle" value="0.5"/>
+ <stat name="FPU_duty_cycle" value="1"/><!-- FPU numbers are already average -->
+ <stat name="ALU_cdb_duty_cycle" value="0.9"/>
+ <stat name="MUL_cdb_duty_cycle" value="0.5"/>
+ <stat name="FPU_cdb_duty_cycle" value="15"/>
+ <component id="system.core0.predictor" name="PBT">
+ <!-- branch predictor; tournament predictor see Alpha implementation -->
+ <param name="local_predictor_size" value="10,3"/>
+ <param name="local_predictor_entries" value="1024"/>
+ <param name="global_predictor_entries" value="4096"/>
+ <param name="global_predictor_bits" value="2"/>
+ <param name="chooser_predictor_entries" value="4096"/>
+ <param name="chooser_predictor_bits" value="2"/>
+ <!-- These parameters can be combined like below in next version
+ <param name="load_predictor" value="10,3,1024"/>
+ <param name="global_predictor" value="4096,2"/>
+ <param name="predictor_chooser" value="4096,2"/>
+ -->
+ </component>
+ <component id="system.core0.itlb" name="itlb">
+ <param name="number_entries" value="1"/>
+ <stat name="total_accesses" value="0"/>
+ <stat name="total_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <!-- there is no write requests to itlb although writes happen to itlb after miss,
+ which is actually a replacement -->
+ </component>
+ <component id="system.core0.icache" name="icache">
+ <!-- there is no write requests to itlb although writes happen to it after miss,
+ which is actually a replacement -->
+ <param name="icache_config" value="16384,128,4,1,1,3,8,0"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <!-- cache_policy;//0 no write or write-though with non-write allocate;1 write-back with write-allocate -->
+ <param name="buffer_sizes" value="16, 16, 16,0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="total_instructions_match_mcpat"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.dtlb" name="dtlb">
+ <param name="number_entries" value="1"/>
+ <stat name="total_accesses" value="0"/>
+ <stat name="total_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.ccache" name="ccache">
+ <!-- all the buffer related are optional -->
+ <param name="ccache_config" value="16384,64,2,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="ccache_read_accesses_match_mcpat"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="ccache_read_misses_match_mcpat"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.tcache" name="tcache">
+ <!-- all the buffer related are optional -->
+ <param name="tcache_config" value="49152,128,8,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="tcache_read_accesses_match_mcpat"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="tcache_read_misses_match_mcpat"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <!--model the shared memory by mimicing dcache-->
+ <component id="system.core0.sharedmemory" name="sharedmemory">
+ <!-- all the buffer related are optional -->
+ <param name="sharedmemory_config" value="98304,16,1,16,1,3,16,0"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="sharedmemory_read_access_match_mcpat"/>
+ <stat name="write_accesses" value="sharedmemory_write_access_match_mcpat"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.dcache" name="dcache">
+ <!-- all the buffer related are optional -->
+ <param name="dcache_config" value="16384,32,4,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="dcache_read_access_match_mcpat"/>
+ <stat name="write_accesses" value="dcache_write_access_match_mcpat"/>
+ <stat name="read_misses" value="dcache_read_miss_match_mcpat"/>
+ <stat name="write_misses" value="dcache_write_miss_match_mcpat"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.BTB" name="BTB">
+ <!-- all the buffer related are optional -->
+ <param name="BTB_config" value="8192,4,2,1, 1,3"/>
+ <!-- the parameters are capacity,block_width,associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ </component>
+ </component>
+ <component id="system.L1Directory0" name="L1Directory0">
+ <param name="Directory_type" value="0"/>
+ <!--0 cam based shadowed tag. 1 directory cache -->
+ <param name="Dir_config" value="2048,1,0,1, 4, 4,8"/>
+ <!-- the parameters are capacity,block_width, associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ <param name="buffer_sizes" value="8, 8, 8, 8"/>
+ <!-- all the buffer related are optional -->
+ <param name="clockrate" value="1400"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw search ports -->
+ <param name="device_type" value="0"/>
+ <!-- altough there are multiple access types,
+ Performance simulator needs to cast them into reads or writes
+ e.g. the invalidates can be considered as writes -->
+ <stat name="read_accesses" value="800000"/>
+ <stat name="write_accesses" value="27276"/>
+ <stat name="read_misses" value="1632"/>
+ <stat name="write_misses" value="183"/>
+ <stat name="conflicts" value="20"/>
+ <stat name="duty_cycle" value="0.45"/>
+ </component>
+ <component id="system.L2Directory0" name="L2Directory0">
+ <param name="Directory_type" value="1"/>
+ <!--0 cam based shadowed tag. 1 directory cache -->
+ <param name="Dir_config" value="1048576,16,16,1,2, 100"/>
+ <!-- the parameters are capacity,block_width, associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ <param name="buffer_sizes" value="8, 8, 8, 8"/>
+ <!-- all the buffer related are optional -->
+ <param name="clockrate" value="1400"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw search ports -->
+ <param name="device_type" value="0"/>
+ <!-- altough there are multiple access types,
+ Performance simulator needs to cast them into reads or writes
+ e.g. the invalidates can be considered as writes -->
+ <stat name="read_accesses" value="0"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.45"/>
+ </component>
+ <component id="system.L20" name="L20">
+ <!-- all the buffer related are optional -->
+ <param name="L2_config" value="131072,128,16,1, 4,23, 64, 1"/>
+ <!-- consider 4-way bank interleaving for Niagara 1 -->
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <param name="clockrate" value="2962"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw ports -->
+ <param name="device_type" value="0"/>
+ <stat name="read_accesses" value="200000"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.5"/>
+ </component>
+
+<!--**********************************************************************-->
+<component id="system.L30" name="L30">
+ <param name="L3_config" value="1048576,64,16,1, 2,100, 64,1"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="clockrate" value="3500"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw ports -->
+ <param name="device_type" value="0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="58824"/>
+ <stat name="write_accesses" value="27276"/>
+ <stat name="read_misses" value="1632"/>
+ <stat name="write_misses" value="183"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.35"/>
+ </component>
+
+
+<!--**********************************************************************-->
+ <component id="system.NoC0" name="noc0">
+ <param name="clockrate" value="700"/>
+ <param name="type" value="1"/>
+ <!-- 1 NoC, O bus -->
+ <param name="horizontal_nodes" value="2"/>
+ <param name="vertical_nodes" value="1"/>
+ <param name="has_global_link" value="0"/>
+ <!-- 1 has global link, 0 does not have global link -->
+ <param name="link_throughput" value="1"/><!--w.r.t clock -->
+ <param name="link_latency" value="1"/><!--w.r.t clock -->
+ <!-- througput >= latency -->
+ <!-- Router architecture -->
+ <param name="input_ports" value="6"/>
+ <param name="output_ports" value="6"/>
+ <param name="virtual_channel_per_port" value="1"/>
+ <!-- input buffer; in classic routers only input ports need buffers -->
+ <param name="flit_bits" value="32"/>
+ <param name="input_buffer_entries_per_vc" value="1"/><!--VCs within the same ports share input buffers whose size is propotional to the number of VCs-->
+ <param name="chip_coverage" value="1"/>
+ <!-- When multiple NOC present, one NOC will cover part of the whole chip. chip_coverage <=1 -->
+ <stat name="total_accesses" value="0"/>
+ <!-- This is the number of total accesses within the whole network not for each router -->
+ <stat name="duty_cycle" value="0.6"/>
+ </component>
+<!--**********************************************************************-->
+<!--**********************************************************************-->
+
+ <component id="system.mem" name="mem">
+ <!-- Main memory property -->
+ <param name="mem_tech_node" value="23"/>
+ <param name="device_clock" value="200"/><!--MHz, this is clock rate of the actual memory device, not the FSB -->
+ <param name="peak_transfer_rate" value="3200"/><!--MB/S-->
+ <param name="internal_prefetch_of_DRAM_chip" value="4"/>
+ <!-- 2 for DDR, 4 for DDR2, 8 for DDR3...-->
+ <!-- the device clock, peak_transfer_rate, and the internal prefetch decide the DIMM property -->
+ <!-- above numbers can be easily found from Wikipedia -->
+ <param name="capacity_per_channel" value="4096"/> <!-- MB -->
+ <!-- capacity_per_Dram_chip=capacity_per_channel/number_of_dimms/number_ranks/Dram_chips_per_rank
+ Current McPAT assumes single DIMMs are used.-->
+ <param name="number_ranks" value="2"/>
+ <param name="num_banks_of_DRAM_chip" value="6"/>
+ <param name="Block_width_of_DRAM_chip" value="64"/> <!-- B -->
+ <param name="output_width_of_DRAM_chip" value="8"/>
+ <!--number of Dram_chips_per_rank=" 72/output_width_of_DRAM_chip-->
+ <!--number of Dram_chips_per_rank=" 72/output_width_of_DRAM_chip-->
+ <param name="page_size_of_DRAM_chip" value="8"/> <!-- 8 or 16 -->
+ <param name="burstlength_of_DRAM_chip" value="8"/>
+ <stat name="memory_accesses" value="1052"/>
+ <stat name="memory_reads" value="1052"/>
+ <stat name="memory_writes" value="1052"/>
+ </component>
+ <component id="system.mc" name="mc">
+ <!-- Memeory controllers are for DDR(2,3...) DIMMs -->
+ <!-- current version of McPAT uses published values for base parameters of memory controller
+ improvments on MC will be added in later versions. -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="mc_clock" value="1848"/><!--DIMM IO bus clock rate MHz DDR2-400 for Niagara 1-->
+ <param name="peak_transfer_rate" value="29568"/><!--MB/S Syed: GTX 470 has 177.4GB/s mem transfer rate with 6 MCs -->
+ <param name="block_size" value="64"/><!--B-->
+ <param name="number_mcs" value="6"/><!-- 6 GDDR5 memory controllers -->
+ <!-- current McPAT only supports homogeneous memory controllers -->
+ <param name="memory_channels_per_mc" value="2"/>
+ <param name="number_ranks" value="1"/>
+ <param name="withPHY" value="0"/>
+ <!-- # of ranks of each channel-->
+ <param name="req_window_size_per_channel" value="16"/>
+ <param name="IO_buffer_size_per_channel" value="16"/>
+ <param name="databus_width" value="32"/>
+ <param name="addressbus_width" value="32"/>
+ <param name="PRT_entries" value="32"/>
+ <!-- # of empirical DRAM model parameter -->
+ <param name="dram_cmd_coeff" value="0"/>
+ <param name="dram_act_coeff" value="0"/>
+ <param name="dram_nop_coeff" value="0"/>
+ <param name="dram_activity_coeff" value="0"/>
+ <param name="dram_pre_coeff" value="3.8475e-8f"/>
+ <param name="dram_rd_coeff" value="7.74707143e-8f"/>
+ <param name="dram_wr_coeff" value="3.54664286e-8f"/>
+ <param name="dram_req_coeff" value="0"/>
+ <param name="dram_const_coeff" value="0"/>
+
+ <!-- McPAT will add the control bus width to the addressbus width automatically -->
+ <stat name="memory_accesses" value="memory_accesses_match_mcpat"/>
+ <stat name="memory_reads" value="memory_reads_match_mcpat"/>
+ <stat name="memory_writes" value="memory_writes_match_mcpat"/>
+ <!-- McPAT does not track individual mc, instead, it takes the total accesses and calculate
+ the average power per MC or per channel. This is sufficent for most application.
+ Further trackdown can be easily added in later versions. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.niu" name="niu">
+ <!-- On chip 10Gb Ethernet NIC, including XAUI Phy and MAC controller -->
+ <!-- For a minimum IP packet size of 84B at 10Gb/s, a new packet arrives every 67.2ns.
+ the low bound of clock rate of a 10Gb MAC is 150Mhz -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="clockrate" value="350"/>
+ <param name="number_units" value="0"/> <!-- unlike PCIe and memory controllers, each Ethernet controller only have one port -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- ratio of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual nic, instead, it takes the total accesses and calculate
+ the average power per nic or per channel. This is sufficent for most application. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.pcie" name="pcie">
+ <!-- On chip PCIe controller, including Phy-->
+ <!-- For a minimum PCIe packet size of 84B at 8Gb/s per lane (PCIe 3.0), a new packet arrives every 84ns.
+ the low bound of clock rate of a PCIe per lane logic is 120Mhz -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="withPHY" value="1"/>
+ <param name="clockrate" value="350"/>
+ <param name="number_units" value="0"/>
+ <param name="num_channels" value="8"/> <!-- 2 ,4 ,8 ,16 ,32 -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- Percentage of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual pcie controllers, instead, it takes the total accesses and calculate
+ the average power per pcie controller or per channel. This is sufficent for most application. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.flashc" name="flashc">
+ <param name="number_flashcs" value="0"/>
+ <param name="type" value="1"/> <!-- 1: low power; 0 high performance -->
+ <param name="withPHY" value="1"/>
+ <param name="peak_transfer_rate" value="200"/><!--Per controller sustainable reak rate MB/S -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- Percentage of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual flash controller, instead, it takes the total accesses and calculate
+ the average power per fc or per channel. This is sufficent for most application -->
+ </component>
+<!--**********************************************************************-->
+
+ </component>
+</component>
\ No newline at end of file diff --git a/configs/tested-cfgs/SM7_GV100/accelwattch_sass_sim.xml b/configs/tested-cfgs/SM7_GV100/accelwattch_sass_sim.xml new file mode 100644 index 0000000..570332d --- /dev/null +++ b/configs/tested-cfgs/SM7_GV100/accelwattch_sass_sim.xml @@ -0,0 +1,613 @@ +<?xml version="1.0" ?>
+<component id="root" name="root">
+ <component id="system" name="system">
+<!-- Copyright (c) 2018-2021, Vijay Kandiah, Junrui Pan, Mahmoud Khairy, Scott Peverelle, Timothy Rogers, Tor M. Aamodt, Nikos Hardavellas
+Northwestern University, Purdue University, The University of British Columbia
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+ list of conditions and the following disclaimer;
+2. Redistributions in binary form must reproduce the above copyright notice,
+ this list of conditions and the following disclaimer in the documentation
+ and/or other materials provided with the distribution;
+3. Neither the names of Northwestern University, Purdue University,
+ The University of British Columbia nor the names of their contributors
+ may be used to endorse or promote products derived from this software
+ without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE. -->
+
+ <!-- DYNAMIC Power Activity Factors -->
+
+ <param name="TOT_INST" value="10" /> <!--Instruction Buffer Activity Factor(Total Warp Instructions)-->
+ <param name="FP_INT" value="4.661" /> <!--Scheduler Activity Factor(Non-Memory Warp Instructons)-->
+ <param name="IC_H" value="8.593489331" /> <!--Instruction Cache Hits -->
+ <param name="IC_M" value="29.735231" /> <!--Instruction Cache Misses-->
+ <param name="DC_RH" value="9.835033124" /> <!--L1 Data Cache Read Hits-->
+ <param name="DC_RM" value="10.95446778" /> <!--L1 Data Cache Read Misses-->
+ <param name="DC_WH" value="0.679656761" /> <!--L1 Data Cache Write Hits-->
+ <param name="DC_WM" value="17.67551799" /> <!--L1 Data Cache Write Misses-->
+ <param name="CC_H" value="0.1107" /> <!--Constant Cache Hits-->
+ <param name="CC_M" value="0.1233" /> <!--Constant Cache Misses-->
+ <param name="SHRD_ACC" value="0.779992642" /> <!--Shared Memory Accesses-->
+ <param name="REG_RD" value="0.100560581" /> <!--Register Reads-->
+ <param name="REG_WR" value="0.140604679" /> <!--Register Writes-->
+ <param name="INT_ACC" value="14.98768151" /> <!--Integer ALU Accesses-->
+ <param name="FP_ACC" value="0.529670751" /> <!--FPU Accesses-->
+ <param name="DP_ACC" value="0.777229051" /> <!--Integer Accesses-->
+ <param name="INT_MUL_ACC" value="0.115098047" /> <!--Integer MUL Accesses-->
+ <param name="FP_MUL_ACC" value="0.089517055" /> <!--FP MUL Accesses-->
+ <param name="FP_SQRT_ACC" value="0.195089274" /> <!--SFU: sqrt Accesses-->
+ <param name="FP_LG_ACC" value="0.125521663" /> <!--SFU: log Accesses-->
+ <param name="FP_SIN_ACC" value="0.13336307" /> <!--SFU: sin/cos Accesses-->
+ <param name="FP_EXP_ACC" value="0.36204415" /> <!--SFU: exponent Accesses-->
+ <param name="DP_MUL_ACC" value="0.1321288" /> <!--DP MUL Accesses-->
+ <param name="TENSOR_ACC" value="0.815454621" /> <!--Tensor Accesses-->
+ <param name="TEX_ACC" value="0.115100088" /> <!--Texture Unit Accesses-->
+ <param name="MEM_RD" value="0.025941068" /> <!--DRAM Reads-->
+ <param name="MEM_WR" value="0.031443719" /> <!--DRAM Writes-->
+ <param name="MEM_PRE" value="0.008647023" /> <!--DRAM Precharge-->
+ <param name="L2_RH" value="1.260867526" /> <!--L2 Data Cache Read Hits-->
+ <param name="L2_RM" value="2.394535301" /> <!--L2 Data Cache Read Misses-->
+ <param name="L2_WH" value="4.124916" /> <!--L2 Data Cache Write Hits-->
+ <param name="L2_WM" value="1.222707601" /> <!--L2 Data Cache Write Misses-->
+ <param name="NOC_A" value="32.09037703" /> <!--Interconnect Network Accesses-->
+ <param name="PIPE_A" value="0.514" /> <!--Pipeline Accesses-->
+
+ <!-- STATIC & CONSTANT Power -->
+
+ <param name="constant_power" value="32.32522272" /> <!--Constant power-->
+ <param name="idle_core_power" value="0.28279166"/> <!--Idle SM power-->
+
+ <param name="static_cat1_flane" value="15.29035866"/> <!--INT (ADD+MUL) First Lane Activation Power-->
+ <param name="static_cat1_addlane" value="0.586233603"/> <!--INT (ADD+MUL) Additional Lane Activation Power-->
+
+ <param name="static_cat2_flane" value="18.6179906"/> <!--INT+FP First Lane Activation Power-->
+ <param name="static_cat2_addlane" value="0.645228013"/> <!--INT+FP Additional Lane Activation Power-->
+
+ <param name="static_cat3_flane" value="19.10017723"/> <!--INT+FP+DP First Lane Activation Power-->
+ <param name="static_cat3_addlane" value="0.726863055"/> <!--INT+FP+DP Additional Lane Activation Power-->
+
+ <param name="static_cat4_flane" value="18.55029744"/> <!--INT+FP+SFU First Lane Activation Power-->
+ <param name="static_cat4_addlane" value="0.6099397"/> <!--INT+FP+SFU Additional Lane Activation Power-->
+
+ <param name="static_cat5_flane" value="14.74826681"/> <!--INT+FP+TEX First Lane Activation Power-->
+ <param name="static_cat5_addlane" value="0.514367937"/> <!--INT+FP+TEX Additional Lane Activation Power-->
+
+ <param name="static_cat6_flane" value="48.94875596"/> <!--INT+FP+TENSOR First Lane Activation Power-->
+ <param name="static_cat6_addlane" value="0.0"/> <!--INT+FP+TENSOR Additional Lane Activation Power-->
+
+ <param name="static_light_flane" value="1.965373811"/> <!--LIGHT_SM First Lane Activation Power-->
+ <param name="static_light_addlane" value="0.003966868"/> <!--LIGHT_SM Additional Lane Activation Power-->
+
+ <param name="static_intadd_flane" value="19.70468506"/> <!--INT ADD First Lane Activation Power-->
+ <param name="static_intadd_addlane" value="0.388578623"/> <!--INT ADD Additional Lane Activation Power-->
+
+ <param name="static_intmul_flane" value="16.64811823"/> <!--INT MUL First Lane Activation Power-->
+ <param name="static_intmul_addlane" value="0.281803166"/> <!--INT MUL Additional Lane Activation Power-->
+
+ <param name="static_geomean_flane" value="17.21745077"/> <!--GEOMEAN First Lane Activation Power-->
+ <param name="static_geomean_addlane" value="0.650630555"/> <!--GEOMEAN Additional Lane Activation Power-->
+ <param name="static_shared_flane" value="31.40965691"/> <!--Shared Memory First Lane Activation Power-->
+ <param name="static_l1_flane" value="34.79491352"/> <!--L1D First Lane Activation Power-->
+ <param name="static_l2_flane" value="17.30654755"/> <!--L2D ADD+MUL First Lane Activation Power-->
+
+
+ <!-- LEAVE BELOW UNCHANGED! These are parameters used for initial GPUWattch model left untouched -->
+
+
+ <param name="GPU_Architecture" value="1"/><!-- 0-G80; 1-Fermi; others not supported -->
+ <param name="architecture" value="1"/> <!-- fermi:1 quadro:2 other: undefined-->
+ <param name="number_of_L1Directories" value="0"/>
+ <param name="number_of_L2Directories" value="0"/>
+ <param name="number_of_L2s" value="1"/> <!-- This number means how many L2 clusters in each cluster there can be multiple banks/ports -->
+ <param name="number_of_L3s" value="0"/> <!-- This number means how many L3 clusters -->
+ <param name="number_of_NoCs" value="1"/>
+ <param name="homogeneous_cores" value="1"/><!--1 means homo -->
+ <param name="homogeneous_L2s" value="1"/>
+ <param name="homogeneous_L1Directorys" value="1"/>
+ <param name="homogeneous_L2Directorys" value="1"/>
+ <param name="homogeneous_L3s" value="1"/>
+ <param name="homogeneous_ccs" value="1"/><!--cache coherece hardware -->
+ <param name="homogeneous_NoCs" value="1"/>
+ <param name="core_tech_node" value="23"/><!-- nm -->
+
+ <param name="number_cache_levels" value="2"/>
+ <param name="interconnect_projection_type" value="0"/><!--0: agressive wire technology; 1: conservative wire technology -->
+ <param name="device_type" value="0"/><!--0: HP(High Performance Type); 1: LSTP(Low standby power) 2: LOP (Low Operating Power) -->
+ <param name="longer_channel_device" value="1"/><!-- 0 no use; 1 use when possible -->
+ <param name="machine_bits" value="32"/>
+ <param name="virtual_address_width" value="32"/>
+ <param name="physical_address_width" value="32"/>
+ <param name="virtual_memory_page_size" value="4096"/>
+
+ <param name="IDLE_CORE_N" value="1"/>
+ <stat name="num_idle_cores" value="0"/><!-- Average Number of idle cores during this period -->
+ <stat name="total_cycles" value="total_cycles_match_mcpat"/>
+ <stat name="idle_cycles" value="idle_cycles_match_mcpat"/>
+ <stat name="busy_cycles" value="busy_cycles_match_mcpat"/>
+ <!--This page size(B) is complete different from the page size in Main memo secction. this page size is the size of
+ virtual memory from OS/Archi perspective; the page size in Main memo secction is the actuall physical line in a DRAM bank -->
+ <!-- *********************** cores ******************* -->
+ <component id="system.core0" name="core0">
+ <!-- Core property -->
+ <param name="clock_rate" value="700"/>
+ <param name="instruction_length" value="32"/>
+ <param name="opcode_width" value="9"/>
+ <!-- address width determins the tag_width in Cache, LSQ and buffers in cache controller
+ default value is machine_bits, if not set -->
+ <param name="machine_type" value="1"/><!-- 1 inorder; 0 OOO-->
+ <!-- inorder/OoO -->
+ <param name="number_hardware_threads" value="32"/>
+ <!-- number_instruction_fetch_ports(icache ports) is always 1 in single-thread processor,
+ it only may be more than one in SMT processors. BTB ports always equals to fetch ports since
+ branch information in consective branch instructions in the same fetch group can be read out from BTB once.-->
+ <param name="fetch_width" value="1"/>
+ <!-- fetch_width determins the size of cachelines of L1 cache block -->
+ <param name="number_instruction_fetch_ports" value="1"/>
+ <param name="decode_width" value="1"/>
+ <!-- decode_width determins the number of ports of the
+ renaming table (both RAM and CAM) scheme -->
+ <param name="issue_width" value="2"/>
+ <!-- issue_width determins the number of ports of Issue window and other logic
+ as in the complexity effective proccessors paper; issue_width==dispatch_width -->
+ <param name="commit_width" value="2"/>
+ <!-- commit_width determins the number of ports of register files -->
+ <param name="fp_issue_width" value="1"/>
+ <param name="prediction_width" value="0"/>
+ <!-- number of branch instructions can be predicted simultannouesl-->
+ <!-- Current version of McPAT does not distinguish int and floating point pipelines
+ Theses parameters are reserved for future use.-->
+ <param name="pipelines_per_core" value="1,1"/>
+ <!--integer_pipeline and floating_pipelines, if the floating_pipelines is 0, then the pipeline is shared-->
+ <param name="pipeline_depth" value="8,8"/>
+ <!-- pipeline depth of int and fp, if pipeline is shared, the second number is the average cycles of fp ops -->
+ <!-- issue and exe unit-->
+ <param name="ALU_per_core" value="32"/>
+ <!-- contains an adder, a shifter, and a logical unit -->
+ <param name="MUL_per_core" value="4"/>
+ <!-- For MUL and Div -->
+ <param name="FPU_per_core" value="32"/>
+ <!-- buffer between IF and ID stage -->
+ <param name="instruction_buffer_size" value="1"/>
+ <!-- buffer between ID and sche/exe stage -->
+ <param name="decoded_stream_buffer_size" value="1"/>
+ <param name="instruction_window_scheme" value="0"/><!-- 0 PHYREG based, 1 RSBASED-->
+ <!-- McPAT support 2 types of OoO cores, RS based and physical reg based-->
+ <param name="instruction_window_size" value="1"/>
+ <param name="fp_instruction_window_size" value="1"/>
+ <!-- the instruction issue Q as in Alpha 21264; The RS as in Intel P6 -->
+ <param name="ROB_size" value="0"/>
+ <!-- each in-flight instruction has an entry in ROB -->
+ <!-- registers -->
+ <!-- SM parameters Added by Syed Gilani -->
+ <param name="rf_banks" value="32"/>
+ <param name="simd_width" value="32"/>
+ <param name="collector_units" value="32"/>
+ <param name="core_clock_ratio" value="2"/>
+ <param name="warp_size" value="32"/>
+
+ <param name="archi_Regs_IRF_size" value="65536"/>
+ <param name="archi_Regs_FRF_size" value="32"/>
+ <!-- if OoO processor, phy_reg number is needed for renaming logic,
+ renaming logic is for both integer and floating point insts. -->
+ <param name="phy_Regs_IRF_size" value="32"/>
+ <param name="phy_Regs_FRF_size" value="32"/>
+ <!-- rename logic -->
+ <param name="rename_scheme" value="0"/>
+ <!-- can be RAM based(0) or CAM based(1) rename scheme
+ RAM-based scheme will have free list, status table;
+ CAM-based scheme have the valid bit in the data field of the CAM
+ both RAM and CAM need RAM-based checkpoint table, checkpoint_depth=# of in_flight instructions;
+ Detailed RAT Implementation see TR -->
+ <param name="register_windows_size" value="0"/>
+ <!-- how many windows in the windowed register file, sun processors;
+ no register windowing is used when this number is 0 -->
+ <!-- In OoO cores, loads and stores can be issued whether inorder(Pentium Pro) or (OoO)out-of-order(Alpha),
+ They will always try to exeute out-of-order though. -->
+ <param name="LSU_order" value="inorder"/>
+ <param name="store_buffer_size" value="32"/>
+ <!-- By default, in-order cores do not have load buffers -->
+ <param name="load_buffer_size" value="32"/>
+ <!-- number of ports refer to sustainable concurrent memory accesses -->
+ <param name="memory_ports" value="2"/>
+ <!-- max_allowed_in_flight_memo_instructions determins the # of ports of load and store buffer
+ as well as the ports of Dcache which is connected to LSU -->
+ <!-- dual-pumped Dcache can be used to save the extra read/write ports -->
+ <param name="RAS_size" value="1"/>
+ <!-- general stats, defines simulation periods;require total, idle, and busy cycles for senity check -->
+ <!-- please note: if target architecture is X86, then all the instrucions refer to (fused) micro-ops -->
+ <stat name="total_instructions" value="total_instructions_match_mcpat"/>
+ <stat name="int_instructions" value="int_instruction_match_mcpat"/>
+ <stat name="fp_instructions" value="flt_instruction_match_mcpat"/>
+ <stat name="branch_instructions" value="branch_instruction_match_mcpat"/>
+ <stat name="branch_mispredictions" value="0"/>
+ <stat name="load_instructions" value="load_instruction_match_mcpat"/>
+ <stat name="store_instructions" value="store_instruction_match_mcpat"/>
+ <stat name="committed_instructions" value="total_instructions_match_mcpat"/>
+ <stat name="committed_int_instructions" value="int_instruction_match_mcpat"/>
+ <stat name="committed_fp_instructions" value="flt_instruction_match_mcpat"/>
+ <stat name="pipeline_duty_cycle" value="0.6"/><!--<=1, runtime_ipc/peak_ipc; averaged for all cores if homogenous -->
+ <!-- the following cycle stats are used for heterogeneouse cores only,
+ please ignore them if homogeneouse cores -->
+ <stat name="total_cycles" value="total_cycles_match_mcpat"/>
+ <stat name="idle_cycles" value="idle_cycles_match_mcpat"/>
+ <stat name="busy_cycles" value="busy_cycles_match_mcpat"/>
+ <!-- instruction buffer stats -->
+ <!-- ROB stats, both RS and Phy based OoOs have ROB
+ performance simulator should capture the difference on accesses,
+ otherwise, McPAT has to guess based on number of commited instructions. -->
+ <stat name="ROB_reads" value="263886"/>
+ <stat name="ROB_writes" value="263886"/>
+ <!-- RAT accesses -->
+ <stat name="rename_accesses" value="263886"/>
+ <stat name="fp_rename_accesses" value="263886"/>
+ <!-- decode and rename stage use this, should be total ic - nop -->
+ <!-- Inst window stats -->
+ <stat name="inst_window_reads" value="263886"/>
+ <stat name="inst_window_writes" value="263886"/>
+ <stat name="inst_window_wakeup_accesses" value="263886"/>
+ <stat name="fp_inst_window_reads" value="263886"/>
+ <stat name="fp_inst_window_writes" value="263886"/>
+ <stat name="fp_inst_window_wakeup_accesses" value="263886"/>
+ <!-- RF accesses -->
+ <stat name="int_regfile_reads" value="int_register_read_access_match_mcpat"/>
+ <stat name="float_regfile_reads" value="int_register_write_access_match_mcpat"/>
+ <stat name="int_regfile_writes" value="float_register_read_access_match_mcpat"/>
+ <stat name="float_regfile_writes" value="float_register_write_access_match_mcpat"/>
+
+ <!-- The following stat is for operand collector power - Added by Syed -->
+ <stat name="non_rf_operands" value="0"/>
+
+ <!-- accesses to the working reg -->
+ <stat name="function_calls" value="0"/>
+ <stat name="context_switches" value="0"/> <!--not used in the McPAT -->
+ <!-- Number of Windowes switches (number of function calls and returns)-->
+ <!-- Alu stats by default, the processor has one FPU that includes the divider and
+ multiplier. The fpu accesses should include accesses to multiplier and divider -->
+ <stat name="ialu_accesses" value="ialu_accesses_match_mcpat"/>
+ <stat name="fpu_accesses" value="fpu_accesses_match_mcpat"/>
+ <stat name="mul_accesses" value="mul_accesses_match_mcpat"/>
+ <stat name="cdb_alu_accesses" value="0"/>
+ <stat name="cdb_mul_accesses" value="0"/>
+ <stat name="cdb_fpu_accesses" value="0"/>
+ <!-- multiple cycle accesses should be counted multiple times,
+ otherwise, McPAT can use internal counter for different floating point instructions
+ to get final accesses. But that needs detailed info for floating point inst mix -->
+ <!-- currently the performance simulator should
+ make sure all the numbers are final numbers,
+ including the explicit read/write accesses,
+ and the implicite accesses such as replacements and etc.
+ Future versions of McPAT may be able to reason the implicite access
+ based on param and stats of last level cache
+ The same rule applies to all cache access stats too! -->
+ <!-- following is AF for max power computation.
+ Do not change them, unless you understand them-->
+ <stat name="IFU_duty_cycle" value="0.25"/>
+ <stat name="LSU_duty_cycle" value="0.25"/>
+ <stat name="MemManU_I_duty_cycle" value="1"/>
+ <stat name="MemManU_D_duty_cycle" value="0.25"/>
+ <stat name="ALU_duty_cycle" value="0.9"/>
+ <stat name="MUL_duty_cycle" value="0.5"/>
+ <stat name="FPU_duty_cycle" value="1"/><!-- FPU numbers are already average -->
+ <stat name="ALU_cdb_duty_cycle" value="0.9"/>
+ <stat name="MUL_cdb_duty_cycle" value="0.5"/>
+ <stat name="FPU_cdb_duty_cycle" value="15"/>
+ <component id="system.core0.predictor" name="PBT">
+ <!-- branch predictor; tournament predictor see Alpha implementation -->
+ <param name="local_predictor_size" value="10,3"/>
+ <param name="local_predictor_entries" value="1024"/>
+ <param name="global_predictor_entries" value="4096"/>
+ <param name="global_predictor_bits" value="2"/>
+ <param name="chooser_predictor_entries" value="4096"/>
+ <param name="chooser_predictor_bits" value="2"/>
+ <!-- These parameters can be combined like below in next version
+ <param name="load_predictor" value="10,3,1024"/>
+ <param name="global_predictor" value="4096,2"/>
+ <param name="predictor_chooser" value="4096,2"/>
+ -->
+ </component>
+ <component id="system.core0.itlb" name="itlb">
+ <param name="number_entries" value="1"/>
+ <stat name="total_accesses" value="0"/>
+ <stat name="total_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <!-- there is no write requests to itlb although writes happen to itlb after miss,
+ which is actually a replacement -->
+ </component>
+ <component id="system.core0.icache" name="icache">
+ <!-- there is no write requests to itlb although writes happen to it after miss,
+ which is actually a replacement -->
+ <param name="icache_config" value="16384,128,4,1,1,3,8,0"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <!-- cache_policy;//0 no write or write-though with non-write allocate;1 write-back with write-allocate -->
+ <param name="buffer_sizes" value="16, 16, 16,0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="total_instructions_match_mcpat"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.dtlb" name="dtlb">
+ <param name="number_entries" value="1"/>
+ <stat name="total_accesses" value="0"/>
+ <stat name="total_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.ccache" name="ccache">
+ <!-- all the buffer related are optional -->
+ <param name="ccache_config" value="16384,64,2,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="ccache_read_accesses_match_mcpat"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="ccache_read_misses_match_mcpat"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.tcache" name="tcache">
+ <!-- all the buffer related are optional -->
+ <param name="tcache_config" value="49152,128,8,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="tcache_read_accesses_match_mcpat"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="tcache_read_misses_match_mcpat"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <!--model the shared memory by mimicing dcache-->
+ <component id="system.core0.sharedmemory" name="sharedmemory">
+ <!-- all the buffer related are optional -->
+ <param name="sharedmemory_config" value="98304,16,1,16,1,3,16,0"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="sharedmemory_read_access_match_mcpat"/>
+ <stat name="write_accesses" value="sharedmemory_write_access_match_mcpat"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.dcache" name="dcache">
+ <!-- all the buffer related are optional -->
+ <param name="dcache_config" value="16384,32,4,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="dcache_read_access_match_mcpat"/>
+ <stat name="write_accesses" value="dcache_write_access_match_mcpat"/>
+ <stat name="read_misses" value="dcache_read_miss_match_mcpat"/>
+ <stat name="write_misses" value="dcache_write_miss_match_mcpat"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.BTB" name="BTB">
+ <!-- all the buffer related are optional -->
+ <param name="BTB_config" value="8192,4,2,1, 1,3"/>
+ <!-- the parameters are capacity,block_width,associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ </component>
+ </component>
+ <component id="system.L1Directory0" name="L1Directory0">
+ <param name="Directory_type" value="0"/>
+ <!--0 cam based shadowed tag. 1 directory cache -->
+ <param name="Dir_config" value="2048,1,0,1, 4, 4,8"/>
+ <!-- the parameters are capacity,block_width, associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ <param name="buffer_sizes" value="8, 8, 8, 8"/>
+ <!-- all the buffer related are optional -->
+ <param name="clockrate" value="1400"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw search ports -->
+ <param name="device_type" value="0"/>
+ <!-- altough there are multiple access types,
+ Performance simulator needs to cast them into reads or writes
+ e.g. the invalidates can be considered as writes -->
+ <stat name="read_accesses" value="800000"/>
+ <stat name="write_accesses" value="27276"/>
+ <stat name="read_misses" value="1632"/>
+ <stat name="write_misses" value="183"/>
+ <stat name="conflicts" value="20"/>
+ <stat name="duty_cycle" value="0.45"/>
+ </component>
+ <component id="system.L2Directory0" name="L2Directory0">
+ <param name="Directory_type" value="1"/>
+ <!--0 cam based shadowed tag. 1 directory cache -->
+ <param name="Dir_config" value="1048576,16,16,1,2, 100"/>
+ <!-- the parameters are capacity,block_width, associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ <param name="buffer_sizes" value="8, 8, 8, 8"/>
+ <!-- all the buffer related are optional -->
+ <param name="clockrate" value="1400"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw search ports -->
+ <param name="device_type" value="0"/>
+ <!-- altough there are multiple access types,
+ Performance simulator needs to cast them into reads or writes
+ e.g. the invalidates can be considered as writes -->
+ <stat name="read_accesses" value="0"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.45"/>
+ </component>
+ <component id="system.L20" name="L20">
+ <!-- all the buffer related are optional -->
+ <param name="L2_config" value="131072,128,16,1, 4,23, 64, 1"/>
+ <!-- consider 4-way bank interleaving for Niagara 1 -->
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <param name="clockrate" value="2962"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw ports -->
+ <param name="device_type" value="0"/>
+ <stat name="read_accesses" value="200000"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.5"/>
+ </component>
+
+<!--**********************************************************************-->
+<component id="system.L30" name="L30">
+ <param name="L3_config" value="1048576,64,16,1, 2,100, 64,1"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="clockrate" value="3500"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw ports -->
+ <param name="device_type" value="0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="58824"/>
+ <stat name="write_accesses" value="27276"/>
+ <stat name="read_misses" value="1632"/>
+ <stat name="write_misses" value="183"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.35"/>
+ </component>
+
+
+<!--**********************************************************************-->
+ <component id="system.NoC0" name="noc0">
+ <param name="clockrate" value="700"/>
+ <param name="type" value="1"/>
+ <!-- 1 NoC, O bus -->
+ <param name="horizontal_nodes" value="2"/>
+ <param name="vertical_nodes" value="1"/>
+ <param name="has_global_link" value="0"/>
+ <!-- 1 has global link, 0 does not have global link -->
+ <param name="link_throughput" value="1"/><!--w.r.t clock -->
+ <param name="link_latency" value="1"/><!--w.r.t clock -->
+ <!-- througput >= latency -->
+ <!-- Router architecture -->
+ <param name="input_ports" value="6"/>
+ <param name="output_ports" value="6"/>
+ <param name="virtual_channel_per_port" value="1"/>
+ <!-- input buffer; in classic routers only input ports need buffers -->
+ <param name="flit_bits" value="32"/>
+ <param name="input_buffer_entries_per_vc" value="1"/><!--VCs within the same ports share input buffers whose size is propotional to the number of VCs-->
+ <param name="chip_coverage" value="1"/>
+ <!-- When multiple NOC present, one NOC will cover part of the whole chip. chip_coverage <=1 -->
+ <stat name="total_accesses" value="0"/>
+ <!-- This is the number of total accesses within the whole network not for each router -->
+ <stat name="duty_cycle" value="0.6"/>
+ </component>
+<!--**********************************************************************-->
+<!--**********************************************************************-->
+
+ <component id="system.mem" name="mem">
+ <!-- Main memory property -->
+ <param name="mem_tech_node" value="23"/>
+ <param name="device_clock" value="200"/><!--MHz, this is clock rate of the actual memory device, not the FSB -->
+ <param name="peak_transfer_rate" value="3200"/><!--MB/S-->
+ <param name="internal_prefetch_of_DRAM_chip" value="4"/>
+ <!-- 2 for DDR, 4 for DDR2, 8 for DDR3...-->
+ <!-- the device clock, peak_transfer_rate, and the internal prefetch decide the DIMM property -->
+ <!-- above numbers can be easily found from Wikipedia -->
+ <param name="capacity_per_channel" value="4096"/> <!-- MB -->
+ <!-- capacity_per_Dram_chip=capacity_per_channel/number_of_dimms/number_ranks/Dram_chips_per_rank
+ Current McPAT assumes single DIMMs are used.-->
+ <param name="number_ranks" value="2"/>
+ <param name="num_banks_of_DRAM_chip" value="6"/>
+ <param name="Block_width_of_DRAM_chip" value="64"/> <!-- B -->
+ <param name="output_width_of_DRAM_chip" value="8"/>
+ <!--number of Dram_chips_per_rank=" 72/output_width_of_DRAM_chip-->
+ <!--number of Dram_chips_per_rank=" 72/output_width_of_DRAM_chip-->
+ <param name="page_size_of_DRAM_chip" value="8"/> <!-- 8 or 16 -->
+ <param name="burstlength_of_DRAM_chip" value="8"/>
+ <stat name="memory_accesses" value="1052"/>
+ <stat name="memory_reads" value="1052"/>
+ <stat name="memory_writes" value="1052"/>
+ </component>
+ <component id="system.mc" name="mc">
+ <!-- Memeory controllers are for DDR(2,3...) DIMMs -->
+ <!-- current version of McPAT uses published values for base parameters of memory controller
+ improvments on MC will be added in later versions. -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="mc_clock" value="1848"/><!--DIMM IO bus clock rate MHz DDR2-400 for Niagara 1-->
+ <param name="peak_transfer_rate" value="29568"/><!--MB/S Syed: GTX 470 has 177.4GB/s mem transfer rate with 6 MCs -->
+ <param name="block_size" value="64"/><!--B-->
+ <param name="number_mcs" value="6"/><!-- 6 GDDR5 memory controllers -->
+ <!-- current McPAT only supports homogeneous memory controllers -->
+ <param name="memory_channels_per_mc" value="2"/>
+ <param name="number_ranks" value="1"/>
+ <param name="withPHY" value="0"/>
+ <!-- # of ranks of each channel-->
+ <param name="req_window_size_per_channel" value="16"/>
+ <param name="IO_buffer_size_per_channel" value="16"/>
+ <param name="databus_width" value="32"/>
+ <param name="addressbus_width" value="32"/>
+ <param name="PRT_entries" value="32"/>
+ <!-- # of empirical DRAM model parameter -->
+ <param name="dram_cmd_coeff" value="0"/>
+ <param name="dram_act_coeff" value="0"/>
+ <param name="dram_nop_coeff" value="0"/>
+ <param name="dram_activity_coeff" value="0"/>
+ <param name="dram_pre_coeff" value="3.8475e-8f"/>
+ <param name="dram_rd_coeff" value="7.74707143e-8f"/>
+ <param name="dram_wr_coeff" value="3.54664286e-8f"/>
+ <param name="dram_req_coeff" value="0"/>
+ <param name="dram_const_coeff" value="0"/>
+
+ <!-- McPAT will add the control bus width to the addressbus width automatically -->
+ <stat name="memory_accesses" value="memory_accesses_match_mcpat"/>
+ <stat name="memory_reads" value="memory_reads_match_mcpat"/>
+ <stat name="memory_writes" value="memory_writes_match_mcpat"/>
+ <!-- McPAT does not track individual mc, instead, it takes the total accesses and calculate
+ the average power per MC or per channel. This is sufficent for most application.
+ Further trackdown can be easily added in later versions. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.niu" name="niu">
+ <!-- On chip 10Gb Ethernet NIC, including XAUI Phy and MAC controller -->
+ <!-- For a minimum IP packet size of 84B at 10Gb/s, a new packet arrives every 67.2ns.
+ the low bound of clock rate of a 10Gb MAC is 150Mhz -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="clockrate" value="350"/>
+ <param name="number_units" value="0"/> <!-- unlike PCIe and memory controllers, each Ethernet controller only have one port -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- ratio of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual nic, instead, it takes the total accesses and calculate
+ the average power per nic or per channel. This is sufficent for most application. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.pcie" name="pcie">
+ <!-- On chip PCIe controller, including Phy-->
+ <!-- For a minimum PCIe packet size of 84B at 8Gb/s per lane (PCIe 3.0), a new packet arrives every 84ns.
+ the low bound of clock rate of a PCIe per lane logic is 120Mhz -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="withPHY" value="1"/>
+ <param name="clockrate" value="350"/>
+ <param name="number_units" value="0"/>
+ <param name="num_channels" value="8"/> <!-- 2 ,4 ,8 ,16 ,32 -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- Percentage of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual pcie controllers, instead, it takes the total accesses and calculate
+ the average power per pcie controller or per channel. This is sufficent for most application. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.flashc" name="flashc">
+ <param name="number_flashcs" value="0"/>
+ <param name="type" value="1"/> <!-- 1: low power; 0 high performance -->
+ <param name="withPHY" value="1"/>
+ <param name="peak_transfer_rate" value="200"/><!--Per controller sustainable reak rate MB/S -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- Percentage of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual flash controller, instead, it takes the total accesses and calculate
+ the average power per fc or per channel. This is sufficent for most application -->
+ </component>
+<!--**********************************************************************-->
+
+ </component>
+</component>
\ No newline at end of file diff --git a/configs/tested-cfgs/SM7_GV100/accelwattch_sass_sim_alt.xml b/configs/tested-cfgs/SM7_GV100/accelwattch_sass_sim_alt.xml new file mode 100644 index 0000000..9998e96 --- /dev/null +++ b/configs/tested-cfgs/SM7_GV100/accelwattch_sass_sim_alt.xml @@ -0,0 +1,613 @@ +<?xml version="1.0" ?>
+<component id="root" name="root">
+ <component id="system" name="system">
+<!-- Copyright (c) 2018-2021, Vijay Kandiah, Junrui Pan, Mahmoud Khairy, Scott Peverelle, Timothy Rogers, Tor M. Aamodt, Nikos Hardavellas
+Northwestern University, Purdue University, The University of British Columbia
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+ list of conditions and the following disclaimer;
+2. Redistributions in binary form must reproduce the above copyright notice,
+ this list of conditions and the following disclaimer in the documentation
+ and/or other materials provided with the distribution;
+3. Neither the names of Northwestern University, Purdue University,
+ The University of British Columbia nor the names of their contributors
+ may be used to endorse or promote products derived from this software
+ without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE. -->
+
+ <!-- DYNAMIC Power Activity Factors -->
+
+ <param name="TOT_INST" value="16" /> <!--Instruction Buffer Activity Factor(Total Warp Instructions)-->
+ <param name="FP_INT" value="2" /> <!--Scheduler Activity Factor(Non-Memory Warp Instructons)-->
+ <param name="IC_H" value="9.967097889" /> <!--Instruction Cache Hits -->
+ <param name="IC_M" value="9.967097889" /> <!--Instruction Cache Misses-->
+ <param name="DC_RH" value="9.382262155" /> <!--L1 Data Cache Read Hits-->
+ <param name="DC_RM" value="9.382262155" /> <!--L1 Data Cache Read Misses-->
+ <param name="DC_WH" value="9.382262155" /> <!--L1 Data Cache Write Hits-->
+ <param name="DC_WM" value="9.382262155" /> <!--L1 Data Cache Write Misses-->
+ <param name="CC_H" value="0.5" /> <!--Constant Cache Hits-->
+ <param name="CC_M" value="0.5" /> <!--Constant Cache Misses-->
+ <param name="SHRD_ACC" value="1.045696662" /> <!--Shared Memory Accesses-->
+ <param name="REG_RD" value="0.096144789" /> <!--Register Reads-->
+ <param name="REG_WR" value="0.134430479" /> <!--Register Writes-->
+ <param name="INT_ACC" value="18.18093465" /> <!--Integer ALU Accesses-->
+ <param name="FP_ACC" value="0.810674358" /> <!--FPU Accesses-->
+ <param name="DP_ACC" value="1.061592439" /> <!--Integer Accesses-->
+ <param name="INT_MUL_ACC" value="0.114254558" /> <!--Integer MUL Accesses-->
+ <param name="FP_MUL_ACC" value="0.155989373" /> <!--FP MUL Accesses-->
+ <param name="FP_SQRT_ACC" value="0.377990392" /> <!--SFU: sqrt Accesses-->
+ <param name="FP_LG_ACC" value="0.155990168" /> <!--SFU: log Accesses-->
+ <param name="FP_SIN_ACC" value="0.155991536" /> <!--SFU: sin/cos Accesses-->
+ <param name="FP_EXP_ACC" value="0.571396061" /> <!--SFU: exponent Accesses-->
+ <param name="DP_MUL_ACC" value="0.155989896" /> <!--DP MUL Accesses-->
+ <param name="TENSOR_ACC" value="1.013271823" /> <!--Tensor Accesses-->
+ <param name="TEX_ACC" value="0.167780179" /> <!--Texture Unit Accesses-->
+ <param name="MEM_RD" value="0.032236806" /> <!--DRAM Reads-->
+ <param name="MEM_WR" value="0.032236806" /> <!--DRAM Writes-->
+ <param name="MEM_PRE" value="0.032236806" /> <!--DRAM Precharge-->
+ <param name="L2_RH" value="1.175516289" /> <!--L2 Data Cache Read Hits-->
+ <param name="L2_RM" value="1.175516289" /> <!--L2 Data Cache Read Misses-->
+ <param name="L2_WH" value="1.175516289" /> <!--L2 Data Cache Write Hits-->
+ <param name="L2_WM" value="1.175516289" /> <!--L2 Data Cache Write Misses-->
+ <param name="NOC_A" value="77.25" /> <!--Interconnect Network Accesses-->
+ <param name="PIPE_A" value="0.4" /> <!--Pipeline Accesses-->
+
+ <!-- STATIC & CONSTANT Power -->
+
+ <param name="constant_power" value="32.32522272" /> <!--Constant power-->
+ <param name="idle_core_power" value="0.28279166"/> <!--Idle SM power-->
+
+ <param name="static_cat1_flane" value="15.29035866"/> <!--INT (ADD+MUL) First Lane Activation Power-->
+ <param name="static_cat1_addlane" value="0.586233603"/> <!--INT (ADD+MUL) Additional Lane Activation Power-->
+
+ <param name="static_cat2_flane" value="18.6179906"/> <!--INT+FP First Lane Activation Power-->
+ <param name="static_cat2_addlane" value="0.645228013"/> <!--INT+FP Additional Lane Activation Power-->
+
+ <param name="static_cat3_flane" value="19.10017723"/> <!--INT+FP+DP First Lane Activation Power-->
+ <param name="static_cat3_addlane" value="0.726863055"/> <!--INT+FP+DP Additional Lane Activation Power-->
+
+ <param name="static_cat4_flane" value="18.55029744"/> <!--INT+FP+SFU First Lane Activation Power-->
+ <param name="static_cat4_addlane" value="0.6099397"/> <!--INT+FP+SFU Additional Lane Activation Power-->
+
+ <param name="static_cat5_flane" value="14.74826681"/> <!--INT+FP+TEX First Lane Activation Power-->
+ <param name="static_cat5_addlane" value="0.514367937"/> <!--INT+FP+TEX Additional Lane Activation Power-->
+
+ <param name="static_cat6_flane" value="48.94875596"/> <!--INT+FP+TENSOR First Lane Activation Power-->
+ <param name="static_cat6_addlane" value="0.0"/> <!--INT+FP+TENSOR Additional Lane Activation Power-->
+
+ <param name="static_light_flane" value="1.965373811"/> <!--LIGHT_SM First Lane Activation Power-->
+ <param name="static_light_addlane" value="0.003966868"/> <!--LIGHT_SM Additional Lane Activation Power-->
+
+ <param name="static_intadd_flane" value="19.70468506"/> <!--INT ADD First Lane Activation Power-->
+ <param name="static_intadd_addlane" value="0.388578623"/> <!--INT ADD Additional Lane Activation Power-->
+
+ <param name="static_intmul_flane" value="16.64811823"/> <!--INT MUL First Lane Activation Power-->
+ <param name="static_intmul_addlane" value="0.281803166"/> <!--INT MUL Additional Lane Activation Power-->
+
+ <param name="static_geomean_flane" value="17.21745077"/> <!--GEOMEAN First Lane Activation Power-->
+ <param name="static_geomean_addlane" value="0.650630555"/> <!--GEOMEAN Additional Lane Activation Power-->
+ <param name="static_shared_flane" value="31.40965691"/> <!--Shared Memory First Lane Activation Power-->
+ <param name="static_l1_flane" value="34.79491352"/> <!--L1D First Lane Activation Power-->
+ <param name="static_l2_flane" value="17.30654755"/> <!--L2D ADD+MUL First Lane Activation Power-->
+
+
+ <!-- LEAVE BELOW UNCHANGED! These are parameters used for initial GPUWattch model left untouched -->
+
+
+ <param name="GPU_Architecture" value="1"/><!-- 0-G80; 1-Fermi; others not supported -->
+ <param name="architecture" value="1"/> <!-- fermi:1 quadro:2 other: undefined-->
+ <param name="number_of_L1Directories" value="0"/>
+ <param name="number_of_L2Directories" value="0"/>
+ <param name="number_of_L2s" value="1"/> <!-- This number means how many L2 clusters in each cluster there can be multiple banks/ports -->
+ <param name="number_of_L3s" value="0"/> <!-- This number means how many L3 clusters -->
+ <param name="number_of_NoCs" value="1"/>
+ <param name="homogeneous_cores" value="1"/><!--1 means homo -->
+ <param name="homogeneous_L2s" value="1"/>
+ <param name="homogeneous_L1Directorys" value="1"/>
+ <param name="homogeneous_L2Directorys" value="1"/>
+ <param name="homogeneous_L3s" value="1"/>
+ <param name="homogeneous_ccs" value="1"/><!--cache coherece hardware -->
+ <param name="homogeneous_NoCs" value="1"/>
+ <param name="core_tech_node" value="23"/><!-- nm -->
+
+ <param name="number_cache_levels" value="2"/>
+ <param name="interconnect_projection_type" value="0"/><!--0: agressive wire technology; 1: conservative wire technology -->
+ <param name="device_type" value="0"/><!--0: HP(High Performance Type); 1: LSTP(Low standby power) 2: LOP (Low Operating Power) -->
+ <param name="longer_channel_device" value="1"/><!-- 0 no use; 1 use when possible -->
+ <param name="machine_bits" value="32"/>
+ <param name="virtual_address_width" value="32"/>
+ <param name="physical_address_width" value="32"/>
+ <param name="virtual_memory_page_size" value="4096"/>
+
+ <param name="IDLE_CORE_N" value="1"/>
+ <stat name="num_idle_cores" value="0"/><!-- Average Number of idle cores during this period -->
+ <stat name="total_cycles" value="total_cycles_match_mcpat"/>
+ <stat name="idle_cycles" value="idle_cycles_match_mcpat"/>
+ <stat name="busy_cycles" value="busy_cycles_match_mcpat"/>
+ <!--This page size(B) is complete different from the page size in Main memo secction. this page size is the size of
+ virtual memory from OS/Archi perspective; the page size in Main memo secction is the actuall physical line in a DRAM bank -->
+ <!-- *********************** cores ******************* -->
+ <component id="system.core0" name="core0">
+ <!-- Core property -->
+ <param name="clock_rate" value="700"/>
+ <param name="instruction_length" value="32"/>
+ <param name="opcode_width" value="9"/>
+ <!-- address width determins the tag_width in Cache, LSQ and buffers in cache controller
+ default value is machine_bits, if not set -->
+ <param name="machine_type" value="1"/><!-- 1 inorder; 0 OOO-->
+ <!-- inorder/OoO -->
+ <param name="number_hardware_threads" value="32"/>
+ <!-- number_instruction_fetch_ports(icache ports) is always 1 in single-thread processor,
+ it only may be more than one in SMT processors. BTB ports always equals to fetch ports since
+ branch information in consective branch instructions in the same fetch group can be read out from BTB once.-->
+ <param name="fetch_width" value="1"/>
+ <!-- fetch_width determins the size of cachelines of L1 cache block -->
+ <param name="number_instruction_fetch_ports" value="1"/>
+ <param name="decode_width" value="1"/>
+ <!-- decode_width determins the number of ports of the
+ renaming table (both RAM and CAM) scheme -->
+ <param name="issue_width" value="2"/>
+ <!-- issue_width determins the number of ports of Issue window and other logic
+ as in the complexity effective proccessors paper; issue_width==dispatch_width -->
+ <param name="commit_width" value="2"/>
+ <!-- commit_width determins the number of ports of register files -->
+ <param name="fp_issue_width" value="1"/>
+ <param name="prediction_width" value="0"/>
+ <!-- number of branch instructions can be predicted simultannouesl-->
+ <!-- Current version of McPAT does not distinguish int and floating point pipelines
+ Theses parameters are reserved for future use.-->
+ <param name="pipelines_per_core" value="1,1"/>
+ <!--integer_pipeline and floating_pipelines, if the floating_pipelines is 0, then the pipeline is shared-->
+ <param name="pipeline_depth" value="8,8"/>
+ <!-- pipeline depth of int and fp, if pipeline is shared, the second number is the average cycles of fp ops -->
+ <!-- issue and exe unit-->
+ <param name="ALU_per_core" value="32"/>
+ <!-- contains an adder, a shifter, and a logical unit -->
+ <param name="MUL_per_core" value="4"/>
+ <!-- For MUL and Div -->
+ <param name="FPU_per_core" value="32"/>
+ <!-- buffer between IF and ID stage -->
+ <param name="instruction_buffer_size" value="1"/>
+ <!-- buffer between ID and sche/exe stage -->
+ <param name="decoded_stream_buffer_size" value="1"/>
+ <param name="instruction_window_scheme" value="0"/><!-- 0 PHYREG based, 1 RSBASED-->
+ <!-- McPAT support 2 types of OoO cores, RS based and physical reg based-->
+ <param name="instruction_window_size" value="1"/>
+ <param name="fp_instruction_window_size" value="1"/>
+ <!-- the instruction issue Q as in Alpha 21264; The RS as in Intel P6 -->
+ <param name="ROB_size" value="0"/>
+ <!-- each in-flight instruction has an entry in ROB -->
+ <!-- registers -->
+ <!-- SM parameters Added by Syed Gilani -->
+ <param name="rf_banks" value="32"/>
+ <param name="simd_width" value="32"/>
+ <param name="collector_units" value="32"/>
+ <param name="core_clock_ratio" value="2"/>
+ <param name="warp_size" value="32"/>
+
+ <param name="archi_Regs_IRF_size" value="65536"/>
+ <param name="archi_Regs_FRF_size" value="32"/>
+ <!-- if OoO processor, phy_reg number is needed for renaming logic,
+ renaming logic is for both integer and floating point insts. -->
+ <param name="phy_Regs_IRF_size" value="32"/>
+ <param name="phy_Regs_FRF_size" value="32"/>
+ <!-- rename logic -->
+ <param name="rename_scheme" value="0"/>
+ <!-- can be RAM based(0) or CAM based(1) rename scheme
+ RAM-based scheme will have free list, status table;
+ CAM-based scheme have the valid bit in the data field of the CAM
+ both RAM and CAM need RAM-based checkpoint table, checkpoint_depth=# of in_flight instructions;
+ Detailed RAT Implementation see TR -->
+ <param name="register_windows_size" value="0"/>
+ <!-- how many windows in the windowed register file, sun processors;
+ no register windowing is used when this number is 0 -->
+ <!-- In OoO cores, loads and stores can be issued whether inorder(Pentium Pro) or (OoO)out-of-order(Alpha),
+ They will always try to exeute out-of-order though. -->
+ <param name="LSU_order" value="inorder"/>
+ <param name="store_buffer_size" value="32"/>
+ <!-- By default, in-order cores do not have load buffers -->
+ <param name="load_buffer_size" value="32"/>
+ <!-- number of ports refer to sustainable concurrent memory accesses -->
+ <param name="memory_ports" value="2"/>
+ <!-- max_allowed_in_flight_memo_instructions determins the # of ports of load and store buffer
+ as well as the ports of Dcache which is connected to LSU -->
+ <!-- dual-pumped Dcache can be used to save the extra read/write ports -->
+ <param name="RAS_size" value="1"/>
+ <!-- general stats, defines simulation periods;require total, idle, and busy cycles for senity check -->
+ <!-- please note: if target architecture is X86, then all the instrucions refer to (fused) micro-ops -->
+ <stat name="total_instructions" value="total_instructions_match_mcpat"/>
+ <stat name="int_instructions" value="int_instruction_match_mcpat"/>
+ <stat name="fp_instructions" value="flt_instruction_match_mcpat"/>
+ <stat name="branch_instructions" value="branch_instruction_match_mcpat"/>
+ <stat name="branch_mispredictions" value="0"/>
+ <stat name="load_instructions" value="load_instruction_match_mcpat"/>
+ <stat name="store_instructions" value="store_instruction_match_mcpat"/>
+ <stat name="committed_instructions" value="total_instructions_match_mcpat"/>
+ <stat name="committed_int_instructions" value="int_instruction_match_mcpat"/>
+ <stat name="committed_fp_instructions" value="flt_instruction_match_mcpat"/>
+ <stat name="pipeline_duty_cycle" value="0.6"/><!--<=1, runtime_ipc/peak_ipc; averaged for all cores if homogenous -->
+ <!-- the following cycle stats are used for heterogeneouse cores only,
+ please ignore them if homogeneouse cores -->
+ <stat name="total_cycles" value="total_cycles_match_mcpat"/>
+ <stat name="idle_cycles" value="idle_cycles_match_mcpat"/>
+ <stat name="busy_cycles" value="busy_cycles_match_mcpat"/>
+ <!-- instruction buffer stats -->
+ <!-- ROB stats, both RS and Phy based OoOs have ROB
+ performance simulator should capture the difference on accesses,
+ otherwise, McPAT has to guess based on number of commited instructions. -->
+ <stat name="ROB_reads" value="263886"/>
+ <stat name="ROB_writes" value="263886"/>
+ <!-- RAT accesses -->
+ <stat name="rename_accesses" value="263886"/>
+ <stat name="fp_rename_accesses" value="263886"/>
+ <!-- decode and rename stage use this, should be total ic - nop -->
+ <!-- Inst window stats -->
+ <stat name="inst_window_reads" value="263886"/>
+ <stat name="inst_window_writes" value="263886"/>
+ <stat name="inst_window_wakeup_accesses" value="263886"/>
+ <stat name="fp_inst_window_reads" value="263886"/>
+ <stat name="fp_inst_window_writes" value="263886"/>
+ <stat name="fp_inst_window_wakeup_accesses" value="263886"/>
+ <!-- RF accesses -->
+ <stat name="int_regfile_reads" value="int_register_read_access_match_mcpat"/>
+ <stat name="float_regfile_reads" value="int_register_write_access_match_mcpat"/>
+ <stat name="int_regfile_writes" value="float_register_read_access_match_mcpat"/>
+ <stat name="float_regfile_writes" value="float_register_write_access_match_mcpat"/>
+
+ <!-- The following stat is for operand collector power - Added by Syed -->
+ <stat name="non_rf_operands" value="0"/>
+
+ <!-- accesses to the working reg -->
+ <stat name="function_calls" value="0"/>
+ <stat name="context_switches" value="0"/> <!--not used in the McPAT -->
+ <!-- Number of Windowes switches (number of function calls and returns)-->
+ <!-- Alu stats by default, the processor has one FPU that includes the divider and
+ multiplier. The fpu accesses should include accesses to multiplier and divider -->
+ <stat name="ialu_accesses" value="ialu_accesses_match_mcpat"/>
+ <stat name="fpu_accesses" value="fpu_accesses_match_mcpat"/>
+ <stat name="mul_accesses" value="mul_accesses_match_mcpat"/>
+ <stat name="cdb_alu_accesses" value="0"/>
+ <stat name="cdb_mul_accesses" value="0"/>
+ <stat name="cdb_fpu_accesses" value="0"/>
+ <!-- multiple cycle accesses should be counted multiple times,
+ otherwise, McPAT can use internal counter for different floating point instructions
+ to get final accesses. But that needs detailed info for floating point inst mix -->
+ <!-- currently the performance simulator should
+ make sure all the numbers are final numbers,
+ including the explicit read/write accesses,
+ and the implicite accesses such as replacements and etc.
+ Future versions of McPAT may be able to reason the implicite access
+ based on param and stats of last level cache
+ The same rule applies to all cache access stats too! -->
+ <!-- following is AF for max power computation.
+ Do not change them, unless you understand them-->
+ <stat name="IFU_duty_cycle" value="0.25"/>
+ <stat name="LSU_duty_cycle" value="0.25"/>
+ <stat name="MemManU_I_duty_cycle" value="1"/>
+ <stat name="MemManU_D_duty_cycle" value="0.25"/>
+ <stat name="ALU_duty_cycle" value="0.9"/>
+ <stat name="MUL_duty_cycle" value="0.5"/>
+ <stat name="FPU_duty_cycle" value="1"/><!-- FPU numbers are already average -->
+ <stat name="ALU_cdb_duty_cycle" value="0.9"/>
+ <stat name="MUL_cdb_duty_cycle" value="0.5"/>
+ <stat name="FPU_cdb_duty_cycle" value="15"/>
+ <component id="system.core0.predictor" name="PBT">
+ <!-- branch predictor; tournament predictor see Alpha implementation -->
+ <param name="local_predictor_size" value="10,3"/>
+ <param name="local_predictor_entries" value="1024"/>
+ <param name="global_predictor_entries" value="4096"/>
+ <param name="global_predictor_bits" value="2"/>
+ <param name="chooser_predictor_entries" value="4096"/>
+ <param name="chooser_predictor_bits" value="2"/>
+ <!-- These parameters can be combined like below in next version
+ <param name="load_predictor" value="10,3,1024"/>
+ <param name="global_predictor" value="4096,2"/>
+ <param name="predictor_chooser" value="4096,2"/>
+ -->
+ </component>
+ <component id="system.core0.itlb" name="itlb">
+ <param name="number_entries" value="1"/>
+ <stat name="total_accesses" value="0"/>
+ <stat name="total_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <!-- there is no write requests to itlb although writes happen to itlb after miss,
+ which is actually a replacement -->
+ </component>
+ <component id="system.core0.icache" name="icache">
+ <!-- there is no write requests to itlb although writes happen to it after miss,
+ which is actually a replacement -->
+ <param name="icache_config" value="16384,128,4,1,1,3,8,0"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <!-- cache_policy;//0 no write or write-though with non-write allocate;1 write-back with write-allocate -->
+ <param name="buffer_sizes" value="16, 16, 16,0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="total_instructions_match_mcpat"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.dtlb" name="dtlb">
+ <param name="number_entries" value="1"/>
+ <stat name="total_accesses" value="0"/>
+ <stat name="total_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.ccache" name="ccache">
+ <!-- all the buffer related are optional -->
+ <param name="ccache_config" value="16384,64,2,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="ccache_read_accesses_match_mcpat"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="ccache_read_misses_match_mcpat"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.tcache" name="tcache">
+ <!-- all the buffer related are optional -->
+ <param name="tcache_config" value="49152,128,8,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="tcache_read_accesses_match_mcpat"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="tcache_read_misses_match_mcpat"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <!--model the shared memory by mimicing dcache-->
+ <component id="system.core0.sharedmemory" name="sharedmemory">
+ <!-- all the buffer related are optional -->
+ <param name="sharedmemory_config" value="98304,16,1,16,1,3,16,0"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="sharedmemory_read_access_match_mcpat"/>
+ <stat name="write_accesses" value="sharedmemory_write_access_match_mcpat"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.dcache" name="dcache">
+ <!-- all the buffer related are optional -->
+ <param name="dcache_config" value="16384,32,4,1,1,3,8,0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 0"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="dcache_read_access_match_mcpat"/>
+ <stat name="write_accesses" value="dcache_write_access_match_mcpat"/>
+ <stat name="read_misses" value="dcache_read_miss_match_mcpat"/>
+ <stat name="write_misses" value="dcache_write_miss_match_mcpat"/>
+ <stat name="conflicts" value="0"/>
+ </component>
+ <component id="system.core0.BTB" name="BTB">
+ <!-- all the buffer related are optional -->
+ <param name="BTB_config" value="8192,4,2,1, 1,3"/>
+ <!-- the parameters are capacity,block_width,associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ </component>
+ </component>
+ <component id="system.L1Directory0" name="L1Directory0">
+ <param name="Directory_type" value="0"/>
+ <!--0 cam based shadowed tag. 1 directory cache -->
+ <param name="Dir_config" value="2048,1,0,1, 4, 4,8"/>
+ <!-- the parameters are capacity,block_width, associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ <param name="buffer_sizes" value="8, 8, 8, 8"/>
+ <!-- all the buffer related are optional -->
+ <param name="clockrate" value="1400"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw search ports -->
+ <param name="device_type" value="0"/>
+ <!-- altough there are multiple access types,
+ Performance simulator needs to cast them into reads or writes
+ e.g. the invalidates can be considered as writes -->
+ <stat name="read_accesses" value="800000"/>
+ <stat name="write_accesses" value="27276"/>
+ <stat name="read_misses" value="1632"/>
+ <stat name="write_misses" value="183"/>
+ <stat name="conflicts" value="20"/>
+ <stat name="duty_cycle" value="0.45"/>
+ </component>
+ <component id="system.L2Directory0" name="L2Directory0">
+ <param name="Directory_type" value="1"/>
+ <!--0 cam based shadowed tag. 1 directory cache -->
+ <param name="Dir_config" value="1048576,16,16,1,2, 100"/>
+ <!-- the parameters are capacity,block_width, associativity,bank, throughput w.r.t. core clock, latency w.r.t. core clock,-->
+ <param name="buffer_sizes" value="8, 8, 8, 8"/>
+ <!-- all the buffer related are optional -->
+ <param name="clockrate" value="1400"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw search ports -->
+ <param name="device_type" value="0"/>
+ <!-- altough there are multiple access types,
+ Performance simulator needs to cast them into reads or writes
+ e.g. the invalidates can be considered as writes -->
+ <stat name="read_accesses" value="0"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.45"/>
+ </component>
+ <component id="system.L20" name="L20">
+ <!-- all the buffer related are optional -->
+ <param name="L2_config" value="131072,128,16,1, 4,23, 64, 1"/>
+ <!-- consider 4-way bank interleaving for Niagara 1 -->
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <param name="clockrate" value="2962"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw ports -->
+ <param name="device_type" value="0"/>
+ <stat name="read_accesses" value="200000"/>
+ <stat name="write_accesses" value="0"/>
+ <stat name="read_misses" value="0"/>
+ <stat name="write_misses" value="0"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.5"/>
+ </component>
+
+<!--**********************************************************************-->
+<component id="system.L30" name="L30">
+ <param name="L3_config" value="1048576,64,16,1, 2,100, 64,1"/>
+ <!-- the parameters are capacity,block_width, associativity, bank, throughput w.r.t. core clock, latency w.r.t. core clock,output_width, cache policy -->
+ <param name="clockrate" value="3500"/>
+ <param name="ports" value="1,1,1"/>
+ <!-- number of r, w, and rw ports -->
+ <param name="device_type" value="0"/>
+ <param name="buffer_sizes" value="16, 16, 16, 16"/>
+ <!-- cache controller buffer sizes: miss_buffer_size(MSHR),fill_buffer_size,prefetch_buffer_size,wb_buffer_size-->
+ <stat name="read_accesses" value="58824"/>
+ <stat name="write_accesses" value="27276"/>
+ <stat name="read_misses" value="1632"/>
+ <stat name="write_misses" value="183"/>
+ <stat name="conflicts" value="0"/>
+ <stat name="duty_cycle" value="0.35"/>
+ </component>
+
+
+<!--**********************************************************************-->
+ <component id="system.NoC0" name="noc0">
+ <param name="clockrate" value="700"/>
+ <param name="type" value="1"/>
+ <!-- 1 NoC, O bus -->
+ <param name="horizontal_nodes" value="2"/>
+ <param name="vertical_nodes" value="1"/>
+ <param name="has_global_link" value="0"/>
+ <!-- 1 has global link, 0 does not have global link -->
+ <param name="link_throughput" value="1"/><!--w.r.t clock -->
+ <param name="link_latency" value="1"/><!--w.r.t clock -->
+ <!-- througput >= latency -->
+ <!-- Router architecture -->
+ <param name="input_ports" value="6"/>
+ <param name="output_ports" value="6"/>
+ <param name="virtual_channel_per_port" value="1"/>
+ <!-- input buffer; in classic routers only input ports need buffers -->
+ <param name="flit_bits" value="32"/>
+ <param name="input_buffer_entries_per_vc" value="1"/><!--VCs within the same ports share input buffers whose size is propotional to the number of VCs-->
+ <param name="chip_coverage" value="1"/>
+ <!-- When multiple NOC present, one NOC will cover part of the whole chip. chip_coverage <=1 -->
+ <stat name="total_accesses" value="0"/>
+ <!-- This is the number of total accesses within the whole network not for each router -->
+ <stat name="duty_cycle" value="0.6"/>
+ </component>
+<!--**********************************************************************-->
+<!--**********************************************************************-->
+
+ <component id="system.mem" name="mem">
+ <!-- Main memory property -->
+ <param name="mem_tech_node" value="23"/>
+ <param name="device_clock" value="200"/><!--MHz, this is clock rate of the actual memory device, not the FSB -->
+ <param name="peak_transfer_rate" value="3200"/><!--MB/S-->
+ <param name="internal_prefetch_of_DRAM_chip" value="4"/>
+ <!-- 2 for DDR, 4 for DDR2, 8 for DDR3...-->
+ <!-- the device clock, peak_transfer_rate, and the internal prefetch decide the DIMM property -->
+ <!-- above numbers can be easily found from Wikipedia -->
+ <param name="capacity_per_channel" value="4096"/> <!-- MB -->
+ <!-- capacity_per_Dram_chip=capacity_per_channel/number_of_dimms/number_ranks/Dram_chips_per_rank
+ Current McPAT assumes single DIMMs are used.-->
+ <param name="number_ranks" value="2"/>
+ <param name="num_banks_of_DRAM_chip" value="6"/>
+ <param name="Block_width_of_DRAM_chip" value="64"/> <!-- B -->
+ <param name="output_width_of_DRAM_chip" value="8"/>
+ <!--number of Dram_chips_per_rank=" 72/output_width_of_DRAM_chip-->
+ <!--number of Dram_chips_per_rank=" 72/output_width_of_DRAM_chip-->
+ <param name="page_size_of_DRAM_chip" value="8"/> <!-- 8 or 16 -->
+ <param name="burstlength_of_DRAM_chip" value="8"/>
+ <stat name="memory_accesses" value="1052"/>
+ <stat name="memory_reads" value="1052"/>
+ <stat name="memory_writes" value="1052"/>
+ </component>
+ <component id="system.mc" name="mc">
+ <!-- Memeory controllers are for DDR(2,3...) DIMMs -->
+ <!-- current version of McPAT uses published values for base parameters of memory controller
+ improvments on MC will be added in later versions. -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="mc_clock" value="1848"/><!--DIMM IO bus clock rate MHz DDR2-400 for Niagara 1-->
+ <param name="peak_transfer_rate" value="29568"/><!--MB/S Syed: GTX 470 has 177.4GB/s mem transfer rate with 6 MCs -->
+ <param name="block_size" value="64"/><!--B-->
+ <param name="number_mcs" value="6"/><!-- 6 GDDR5 memory controllers -->
+ <!-- current McPAT only supports homogeneous memory controllers -->
+ <param name="memory_channels_per_mc" value="2"/>
+ <param name="number_ranks" value="1"/>
+ <param name="withPHY" value="0"/>
+ <!-- # of ranks of each channel-->
+ <param name="req_window_size_per_channel" value="16"/>
+ <param name="IO_buffer_size_per_channel" value="16"/>
+ <param name="databus_width" value="32"/>
+ <param name="addressbus_width" value="32"/>
+ <param name="PRT_entries" value="32"/>
+ <!-- # of empirical DRAM model parameter -->
+ <param name="dram_cmd_coeff" value="0"/>
+ <param name="dram_act_coeff" value="0"/>
+ <param name="dram_nop_coeff" value="0"/>
+ <param name="dram_activity_coeff" value="0"/>
+ <param name="dram_pre_coeff" value="3.8475e-8f"/>
+ <param name="dram_rd_coeff" value="7.74707143e-8f"/>
+ <param name="dram_wr_coeff" value="3.54664286e-8f"/>
+ <param name="dram_req_coeff" value="0"/>
+ <param name="dram_const_coeff" value="0"/>
+
+ <!-- McPAT will add the control bus width to the addressbus width automatically -->
+ <stat name="memory_accesses" value="memory_accesses_match_mcpat"/>
+ <stat name="memory_reads" value="memory_reads_match_mcpat"/>
+ <stat name="memory_writes" value="memory_writes_match_mcpat"/>
+ <!-- McPAT does not track individual mc, instead, it takes the total accesses and calculate
+ the average power per MC or per channel. This is sufficent for most application.
+ Further trackdown can be easily added in later versions. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.niu" name="niu">
+ <!-- On chip 10Gb Ethernet NIC, including XAUI Phy and MAC controller -->
+ <!-- For a minimum IP packet size of 84B at 10Gb/s, a new packet arrives every 67.2ns.
+ the low bound of clock rate of a 10Gb MAC is 150Mhz -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="clockrate" value="350"/>
+ <param name="number_units" value="0"/> <!-- unlike PCIe and memory controllers, each Ethernet controller only have one port -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- ratio of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual nic, instead, it takes the total accesses and calculate
+ the average power per nic or per channel. This is sufficent for most application. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.pcie" name="pcie">
+ <!-- On chip PCIe controller, including Phy-->
+ <!-- For a minimum PCIe packet size of 84B at 8Gb/s per lane (PCIe 3.0), a new packet arrives every 84ns.
+ the low bound of clock rate of a PCIe per lane logic is 120Mhz -->
+ <param name="type" value="0"/> <!-- 1: low power; 0 high performance -->
+ <param name="withPHY" value="1"/>
+ <param name="clockrate" value="350"/>
+ <param name="number_units" value="0"/>
+ <param name="num_channels" value="8"/> <!-- 2 ,4 ,8 ,16 ,32 -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- Percentage of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual pcie controllers, instead, it takes the total accesses and calculate
+ the average power per pcie controller or per channel. This is sufficent for most application. -->
+ </component>
+<!--**********************************************************************-->
+ <component id="system.flashc" name="flashc">
+ <param name="number_flashcs" value="0"/>
+ <param name="type" value="1"/> <!-- 1: low power; 0 high performance -->
+ <param name="withPHY" value="1"/>
+ <param name="peak_transfer_rate" value="200"/><!--Per controller sustainable reak rate MB/S -->
+ <stat name="duty_cycle" value="1.0"/> <!-- achievable max load <= 1.0 -->
+ <stat name="total_load_perc" value="0.7"/> <!-- Percentage of total achived load to total achivable bandwidth -->
+ <!-- McPAT does not track individual flash controller, instead, it takes the total accesses and calculate
+ the average power per fc or per channel. This is sufficent for most application -->
+ </component>
+<!--**********************************************************************-->
+
+ </component>
+</component>
\ No newline at end of file diff --git a/configs/tested-cfgs/SM7_GV100/config_volta_islip.icnt b/configs/tested-cfgs/SM7_GV100/config_volta_islip.icnt new file mode 100644 index 0000000..5ad7ecd --- /dev/null +++ b/configs/tested-cfgs/SM7_GV100/config_volta_islip.icnt @@ -0,0 +1,74 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 144; +n = 1; + +// Routing + +routing_function = dest_tag; + + +// Flow control + +num_vcs = 1; +vc_buf_size = 256; +input_buffer_size = 256; +ejection_buffer_size = 256; +boundary_buffer_size = 256; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 2.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/tested-cfgs/SM7_GV100/gpgpusim.config b/configs/tested-cfgs/SM7_GV100/gpgpusim.config new file mode 100644 index 0000000..8d2b101 --- /dev/null +++ b/configs/tested-cfgs/SM7_GV100/gpgpusim.config @@ -0,0 +1,237 @@ +# Copyright (c) 2018-2021, Vijay Kandiah, Junrui Pan, Mahmoud Khairy, Scott Peverelle, Timothy Rogers, Tor M. Aamodt, Nikos Hardavellas +# Northwestern University, Purdue University, The University of British Columbia +# All rights reserved. + +# Redistribution and use in source and binary forms, with or without +# modification, are permitted provided that the following conditions are met: + +# 1. Redistributions of source code must retain the above copyright notice, this +# list of conditions and the following disclaimer; +# 2. Redistributions in binary form must reproduce the above copyright notice, +# this list of conditions and the following disclaimer in the documentation +# and/or other materials provided with the distribution; +# 3. Neither the names of Northwestern University, Purdue University, +# The University of British Columbia nor the names of their contributors +# may be used to endorse or promote products derived from this software +# without specific prior written permission. + +# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +# ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE +# LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +# CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +# SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +# INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +# CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +# ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +# POSSIBILITY OF SUCH DAMAGE. + + +# This config models the Volta Quadro V100 +# For more info about volta architecture: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# Device Limits +-gpgpu_stack_size_limit 1024 +-gpgpu_heap_size_limit 8388608 +-gpgpu_runtime_sync_depth_limit 2 +-gpgpu_runtime_pending_launch_count_limit 2048 +-gpgpu_kernel_launch_latency 5000 +-gpgpu_TB_launch_latency 0 +-gpgpu_max_concurrent_kernel 128 + +# Compute Capability +-gpgpu_compute_capability_major 7 +-gpgpu_compute_capability_minor 0 + +# PTX execution-driven +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 80 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 32 +-gpgpu_n_sub_partition_per_mchannel 2 +-gpgpu_clock_gated_lanes 1 + +# volta clock domains +#-gpgpu_clock_domains <Core Clock>:<Interconnect Clock>:<L2 Clock>:<DRAM Clock> +-gpgpu_clock_domains 1447.0:1447.0:1447.0:850.0 +# boost mode +# -gpgpu_clock_domains 1628.0:1628.0:1628.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 +-gpgpu_registers_per_block 65536 +-gpgpu_occupancy_sm_number 70 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_INT,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_INT,OC_EX_SFU,OC_EX_MEM,EX_WB,ID_OC_TENSOR_CORE,OC_EX_TENSOR_CORE +## Volta GV100 has 4 SP SIMD units, 4 SFU units, 4 DP units per core, 4 Tensor core units +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,4,4,4,4,4,4,4,4,8,4,4 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 +-gpgpu_num_dp_units 4 +-gpgpu_num_int_units 4 +-gpgpu_tensor_core_avail 1 +-gpgpu_num_tensor_core_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +-ptx_opcode_latency_int 4,13,4,5,145,21 +-ptx_opcode_initiation_int 2,2,2,2,8,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 +-ptx_opcode_latency_sfu 100 +-ptx_opcode_initiation_sfu 8 +-ptx_opcode_latency_tesnor 64 +-ptx_opcode_initiation_tensor 64 + +# Volta has sub core model, in which each scheduler has its own register file and EUs +# i.e. schedulers are isolated +-gpgpu_sub_core_model 1 +# disable specialized operand collectors and use generic operand collectors instead +-gpgpu_enable_specialized_operand_collector 0 +-gpgpu_operand_collector_num_units_gen 8 +-gpgpu_operand_collector_num_in_ports_gen 8 +-gpgpu_operand_collector_num_out_ports_gen 8 +# volta has 8 banks, 4 schedulers, two banks per scheduler +# we increase #banks to 16 to mitigate the effect of Regisrer File Cache (RFC) which we do not implement in the current version +-gpgpu_num_reg_banks 16 +-gpgpu_reg_file_port_throughput 2 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 70 + +# Volta has four schedulers per core +-gpgpu_num_sched_per_core 4 +# Greedy then oldest scheduler +-gpgpu_scheduler lrr +## In Volta, a warp scheduler can issue 1 inst per cycle +-gpgpu_max_insn_issue_per_warp 1 +-gpgpu_dual_issue_diff_exec_units 1 + +## L1/shared memory configuration +# <nsets>:<bsize>:<assoc>,<rep>:<wr>:<alloc>:<wr_alloc>:<set_index_fn>,<mshr>:<N>:<merge>,<mq>:**<fifo_entry> +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Defualt config is 32KB DL1 and 96KB shared memory +# In Volta, we assign the remaining shared memory to L1 cache +# if the assigned shd mem = 0, then L1 cache = 128KB +# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x +# disable this mode in case of multi kernels/apps execution +-gpgpu_adaptive_cache_config 1 +-gpgpu_shmem_option 0,8,16,32,64,96 +-gpgpu_unified_l1d_size 128 +# L1 cache configuration +-gpgpu_l1_banks 4 +-gpgpu_cache:dl1 S:4:128:64,L:T:m:L:L,A:512:8,16:0,32 +-gpgpu_l1_cache_write_ratio 25 +-gpgpu_l1_latency 20 +-gpgpu_gmem_skip_L1D 0 +-gpgpu_flush_l1_cache 1 +-gpgpu_n_cluster_ejection_buffer_size 32 +# shared memory configuration +-gpgpu_shmem_size 98304 +-gpgpu_shmem_sizeDefault 98304 +-gpgpu_shmem_per_block 65536 +-gpgpu_smem_latency 20 + +# 32 sets, each 128 bytes 24-way for each memory sub partition (96 KB per memory sub partition). This gives us 6MB L2 cache +-gpgpu_cache:dl2 S:32:128:24,L:B:m:L:P,A:192:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-gpgpu_perf_sim_memcpy 1 +-gpgpu_memory_partition_indexing 2 + +# 128 KB Inst. +-gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 +-gpgpu_inst_fetch_throughput 4 +# 128 KB Tex +# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod +-gpgpu_tex_cache:l1 N:4:128:256,L:R:m:N:L,T:512:8,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 +-gpgpu_perfect_inst_const_cache 1 + +# interconnection +#-network_mode 1 +#-inter_config_file config_volta_islip.icnt +# use built-in local xbar +-network_mode 2 +-icnt_in_buffer_limit 512 +-icnt_out_buffer_limit 512 +-icnt_subnets 2 +-icnt_flit_size 40 +-icnt_arbiter_algo 1 + +# memory partition latency config +-gpgpu_l2_rop_latency 160 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 192 + +# for HBM, three stacks, 24 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCB.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 850 MHZ, V100 HBM runs at 850 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=2:CDLR=3:WR=10:nbkgrp=4:CCDL=2:RTPL=3" + +# HBM has dual bus interface, in which it can issue two col and row commands at a time +-dram_dual_bus_interface 1 +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-dram_seperate_write_queue_enable 1 +#-dram_write_queue_size 64:56:32 + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0
\ No newline at end of file diff --git a/configs/tested-cfgs/SM7_QV100/gpgpusim.config b/configs/tested-cfgs/SM7_QV100/gpgpusim.config index 8d2b101..1b55aaf 100644 --- a/configs/tested-cfgs/SM7_QV100/gpgpusim.config +++ b/configs/tested-cfgs/SM7_QV100/gpgpusim.config @@ -69,7 +69,7 @@ # volta clock domains #-gpgpu_clock_domains <Core Clock>:<Interconnect Clock>:<L2 Clock>:<DRAM Clock> --gpgpu_clock_domains 1447.0:1447.0:1447.0:850.0 +-gpgpu_clock_domains 1132.0:1132.0:1132.0:850.0 # boost mode # -gpgpu_clock_domains 1628.0:1628.0:1628.0:850.0 |