Inquiry about the status of H100 config

[GaoHengSiang]

Hello, thank you for this wonderful tool.

I was trying to simulate the cycle count of an H100 performing GEMM, however I only have access to an A40 GPU.
So I traced gemm (tencore and normal) from Deepbench_nvidia with the A40 GPU, then tried to run it with the H100-SASS config.

The H100 configs I gathered from these posts:

#344
accel-sim/gpgpu-sim_distribution#80

The simulation ultimately failed with SEGF, and I learned from #161 that it was to be expected.

The error log:

gemm_bench-inference_half_10_10_10_0_0--H100-SASS. Status=SEGF
Last 10 line of /home/gaohengsiang/accelsim/accel-sim-framework/util/job_launching/../../sim_run_12.0/gemm_bench/inference_half_10_10_10_0_0/H100-SASS/gemm_bench-inference_half_10_10_10_0_0.accelsim-commit-3c96d32_modified_0.0_25-11-25-22-09-15gpgpu-sim_git-commit-b18ee397_modified_0.0.o150
------------------
thread block = 0,0,1
GPGPU-Sim uArch: Shader 64 bind to kernel 7 '_ZN7cutlass6KernelI58cutlass_80_wmma_tensorop_h161616gemm_16x16_128x2_nn_align2EEvNT_6ParamsE'
thread block = 0,0,0
GPGPU-Sim: Reconfigure L1 cache to 96KB
GPGPU-Sim uArch: Shader 63 bind to kernel 7 '_ZN7cutlass6KernelI58cutlass_80_wmma_tensorop_h161616gemm_16x16_128x2_nn_align2EEvNT_6ParamsE'
launching kernel name: _ZN7cutlass6KernelI58cutlass_80_wmma_tensorop_h161616gemm_16x16_128x2_nn_align2EEvNT_6ParamsE uid: 7 cuda_stream_id: 0
Header info loaded for kernel command : ./traces/kernel-7-ctx_0x6245c0f86880.traceg.xz
-enable lineinfo = 0
-accelsim tracer version = 5
-nvbit version = 1.7.6
------------------

Contents of /home/gaohengsiang/accelsim/accel-sim-framework/util/job_launching/../../sim_run_12.0/gemm_bench/inference_half_10_10_10_0_0/H100-SASS/gemm_bench-inference_half_10_10_10_0_0.accelsim-commit-3c96d32_modified_0.0_25-11-25-22-09-15gpgpu-sim_git-commit-b18ee397_modified_0.0.e150
------------------
/home/gaohengsiang/accelsim/accel-sim-framework/util/job_launching/../../sim_run_12.0/gemm_bench/inference_half_10_10_10_0_0/H100-SASS/slurm.sim: line 54: 3984007 Segmentation fault (core dumped)
/home/gaohengsiang/accelsim/accel-sim-framework/util/job_launching/../../sim_run_12.0/gpgpu-sim-builds/accelsim-commit-3c96d32_modified_0.0_25-11-25-22-09-15gpgpu-sim_git-commit-b18ee397_modified_0.0/accel-sim.out -config ./gpgpusim.config -trace ./traces/kernelslist.g

All jobs seemed to SEGF at kernel-7, and mentioned something related to wmma.

This led me to two questions:

1. Since the h100-test branch was closed, is the H100 config usable at all, or is it currently unsalvageable?
2. Is there a way to approximate H100 performance by modifying the A100 config? (simulation with A100 config ran fine)

[JRPan]

Can you see where the SEGF is coming from? Build it in debug, and run it inside gdb.

Generally, you should not run one config (Hopper) with traces generated from another architecture (Ampere). But it seems like the configuration is wrong, so in this case, it might be fixable.

Also, it would be helpful if you could upload the configs you are using.

[GaoHengSiang]

Here are the config files:

trace.config.txt
config_hopper_islip.icnt.txt
gpgpusim.config.txt
hopper_opcode.h.txt

Had to convert them to .txt because github doesn't support the original file types

I had no experience with gdb, so it will take some time to figure out