From fb871c36a565bff2e25f457b8e7a0d5a6ffc4b7f Mon Sep 17 00:00:00 2001 From: Scott Peverelle Date: Wed, 25 May 2016 11:25:09 -0800 Subject: Tweaked Maxwell config to improve correspondence in a number of areas; clock speed, DRAM bus width, L2 cache size etc. Adjusted parse_and_compare to match new clock frequency as per config file. [git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 21818] --- configs/GeForceGTX750Ti/gpgpusim.config | 27 +++++++++++++-------------- 1 file changed, 13 insertions(+), 14 deletions(-) (limited to 'configs') diff --git a/configs/GeForceGTX750Ti/gpgpusim.config b/configs/GeForceGTX750Ti/gpgpusim.config index 0e69098..9a3d73b 100644 --- a/configs/GeForceGTX750Ti/gpgpusim.config +++ b/configs/GeForceGTX750Ti/gpgpusim.config @@ -18,10 +18,10 @@ #-gpgpu_clock_domains ::: # In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided # by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 --gpgpu_clock_domains 1080.0:1080.0:1080.0:924.0 +-gpgpu_clock_domains 1137.0:1137.0:1137.0:2700.0 # shader core pipeline config --gpgpu_shader_registers 32768 +-gpgpu_shader_registers 65536 # This implies a maximum of 64 warps/SM -gpgpu_shader_core_pipeline 2048:32 @@ -32,31 +32,30 @@ # ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB -gpgpu_pipeline_widths 2,1,1,2,1,1,2 -gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 +-gpgpu_num_sfu_units 32 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,1,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_int 6,12,13,13,210 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 6,12,6,6,374 +-ptx_opcode_initiation_fp 1,1,1,1,4 -ptx_opcode_latency_dp 8,19,8,8,330 -ptx_opcode_initiation_dp 8,16,8,8,130 - # In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb # ::,::::,::,:** # ** Optional parameter - Required when mshr_type==Texture Fifo # Note: Hashing set index function (H) only applies to a set size of 32 or 64. -gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 --gpgpu_shmem_size 49152 +-gpgpu_shmem_size 65536 # The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected #-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 #-gpgpu_shmem_size 16384 -# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache --gpgpu_cache:dl2 64:128:8,L:B:m:W:L,A:32:4,4:0,32 +# 1024 sets, each 128 bytes 8-way for each memory sub partition. This gives 2MB L2 cache +-gpgpu_cache:dl2 1024:128:8,L:B:m:W:L,A:32:4,4:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 @@ -97,7 +96,7 @@ # for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition -gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 +-gpgpu_dram_buswidth 32 -gpgpu_dram_burst_length 8 -dram_data_command_freq_ratio 4 # GDDR5 is QDR -gpgpu_mem_address_mask 1 @@ -108,8 +107,8 @@ -gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 +# Maxwell has four schedulers per core +-gpgpu_num_sched_per_core 4 # Two Level Scheduler with active and pending pools #-gpgpu_scheduler two_level_active:6:0:1 # Loose round robbin scheduler -- cgit v1.3