diff options
| author | negargoli93 <[email protected]> | 2018-06-20 18:20:05 -0700 |
|---|---|---|
| committer | negargoli93 <[email protected]> | 2018-06-20 18:20:05 -0700 |
| commit | 262663ac90d2aa801d6af1eb9bf8a75ee9a5bb18 (patch) | |
| tree | 4d2d668571bc8e62e4c25f2a1fe9f0283ea0e5a5 /cuda-kernels | |
| parent | acc52418fe873712ed56a072190356ecc874a35b (diff) | |
Tensor core timing model
Diffstat (limited to 'cuda-kernels')
| -rwxr-xr-x | cuda-kernels/gpgpusim.config | 7 |
1 files changed, 5 insertions, 2 deletions
diff --git a/cuda-kernels/gpgpusim.config b/cuda-kernels/gpgpusim.config index 306d7f9..69a110f 100755 --- a/cuda-kernels/gpgpusim.config +++ b/cuda-kernels/gpgpusim.config @@ -33,10 +33,10 @@ # ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,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 4,1,1,4,1,1,6 +-gpgpu_pipeline_widths 4,1,1,1,4,1,1,1,,6 -gpgpu_num_sp_units 4 -gpgpu_num_sfu_units 1 - +-gpgpu_num_tensor_core_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 @@ -72,11 +72,14 @@ ## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units -gpgpu_operand_collector_num_units_sp 20 -gpgpu_operand_collector_num_units_sfu 4 +-gpgpu_operand_collector_num_units_tensor_core 24 -gpgpu_operand_collector_num_units_mem 8 -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_tensor_core 1 +-gpgpu_operand_collector_num_out_ports_tensor_core 1 -gpgpu_operand_collector_num_in_ports_mem 1 -gpgpu_operand_collector_num_out_ports_mem 1 # gpgpu_num_reg_banks should be increased to 32, but it gives an error! |