1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
|
# This config models the Pascal GP100
# For more info about this card, see Nvidia White paper
# https://images.nvidia.com/content/pdf/tesla/whitepaper/pascal-architecture-whitepaper.pdf
# functional simulator specification
-gpgpu_ptx_instruction_classification 0
-gpgpu_ptx_sim_mode 0
-gpgpu_ptx_force_max_capability 60
# SASS execution (only supported with CUDA >= 4.0)
-gpgpu_ptx_convert_to_ptxplus 0
-gpgpu_ptx_save_converted_ptxplus 0
# high level architecture configuration
-gpgpu_n_clusters 28
-gpgpu_n_cores_per_cluster 2
-gpgpu_n_mem 32
-gpgpu_n_sub_partition_per_mchannel 1
# Pscal clock domains
#-gpgpu_clock_domains <Core Clock>:<Interconnect Clock>:<L2 Clock>:<DRAM Clock>
# Pascal NVIDIA GP100 clock domains are adopted from
# https://en.wikipedia.org/wiki/Nvidia_Tesla
-gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0
# shader core pipeline config
-gpgpu_shader_registers 65536
# 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_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB
## Pascal GP102 has 4 SP SIMD units and 1 SFU unit
## we need to scale the number of pipeline registers to be equal to the number of SP units
-gpgpu_pipeline_widths 2,2,1,2,2,1,4
-gpgpu_num_sp_units 2
-gpgpu_num_sfu_units 1
# Instruction latencies and initiation intervals
# "ADD,MAX,MUL,MAD,DIV"
# SFU is 32-width in pascal, then dp units initiation is 1 cycle
-ptx_opcode_latency_int 4,13,4,5,145
-ptx_opcode_initiation_int 1,1,1,1,4
-ptx_opcode_latency_fp 4,13,4,5,39
-ptx_opcode_initiation_fp 1,2,1,1,4
-ptx_opcode_latency_dp 8,19,8,8,330
-ptx_opcode_initiation_dp 2,2,2,2,130
# <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
# Note: Hashing set index function (H) only applies to a set size of 32 or 64.
# Pascal GP100 has 64KB Shared memory
# Pascal GP100 has 48KB L1 cache
-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16
-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16
-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16
-gpgpu_shmem_size 65536
-gpgpu_shmem_size_PrefL1 65536
-gpgpu_shmem_size_PrefShared 65536
-gmem_skip_L1D 0
# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache
-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32
-gpgpu_cache:dl2_texture_only 0
-gpgpu_dram_partition_queues 32:32:32:32
# 4 KB Inst.
-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4
# 48 KB Tex
-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2
# 12 KB Const
-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4
# enable operand collector
-gpgpu_operand_collector_num_units_sp 14
-gpgpu_operand_collector_num_units_sfu 8
-gpgpu_operand_collector_num_units_mem 10
-gpgpu_operand_collector_num_in_ports_sp 4
-gpgpu_operand_collector_num_out_ports_sp 4
-gpgpu_operand_collector_num_in_ports_sfu 1
-gpgpu_operand_collector_num_out_ports_sfu 1
-gpgpu_operand_collector_num_in_ports_mem 1
-gpgpu_operand_collector_num_out_ports_mem 1
-gpgpu_num_reg_banks 32
# shared memory bankconflict detection
-gpgpu_shmem_num_banks 32
-gpgpu_shmem_limited_broadcast 0
-gpgpu_shmem_warp_parts 1
## In Pascal, a warp scheduler can issue 2 insts per cycle
-gpgpu_max_insn_issue_per_warp 2
# interconnection
-network_mode 1
-inter_config_file config_fermi_islip.icnt
# memory partition latency config
-rop_latency 120
-dram_latency 100
# dram model config
-gpgpu_dram_scheduler 1
# The DRAM return queue and the scheduler queue together should provide buffer
# to sustain the memory level parallelism to tolerate DRAM latency
# To allow 100% DRAM utility, there should at least be enough buffer to sustain
# the minimum DRAM latency (100 core cycles). I.e.
# Total buffer space required = 100 x 924MHz / 700MHz = 132
-gpgpu_frfcfs_dram_sched_queue_size 64
-gpgpu_dram_return_queue_size 116
# for HBM, 32 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.RBBBBCCC.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 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ
-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34:
CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3"
# Fermi has two schedulers per core
-gpgpu_num_sched_per_core 2
# Two Level Scheduler with active and pending pools
#-gpgpu_scheduler two_level_active:6:0:1
# Loose round robbin scheduler
#-gpgpu_scheduler lrr
# Greedy then oldest scheduler
-gpgpu_scheduler gto
# stat collection
-gpgpu_memlatency_stat 14
-gpgpu_runtime_stat 500
-enable_ptx_file_line_stats 1
-visualizer_enabled 0
# power model configs
-power_simulation_enabled 0
-gpuwattch_xml_file gpuwattch_gtx480.xml
# tracing functionality
#-trace_enabled 1
#-trace_components WARP_SCHEDULER,SCOREBOARD
#-trace_sampling_core 0
|
