mmapeak
More actions
Reported 8 GB and 10 GB
mmapeak is a small CUDA benchmark that measures peak tensor-core throughput for each data type. In the Discord it is the tool to ask a seller to run before you buy, to show the card's tensor cores work at full rate; a PyTorch GEMM benchmark is the second option. Source: @jimtjames, 2026-10-11 Author: GitHub ReinForce-II. Repository: https://github.com/ReinForce-II/mmapeak
Build and run
git clone https://github.com/ReinForce-II/mmapeak && cd mmapeak
mkdir build && cd build && cmake .. && make
./mmapeak # default target 3 s per kernel
./mmapeak -t 10 # longer per kernel
Each kernel prints run: <ms> <T(fl)ops> or not supported. The CMake file builds for sm_75 to sm_120; CUDA 12.8.1 or later is recommended. Source: repo github.com/ReinForce-II/mmapeak by ReinForce-II on GitHub, posted by @jimtjames, 2026-10-11.
Reading it on a CMP 170HX
The 170HX is sm_80 (GA100). Only the INT4, INT8, FP16, BF16 and TF32 rows run; FP8, FP6, FP4 and MX rows print not supported. The rows that matter: mma_s4s4s32_8_8_32, mma_s8s8s32_*, mma_f16f16f16_*, mma_f16f16f32_*, mma_bf16bf16f32_*, mma_tf32tf32f32_16_16_8. Source: repo github.com/ReinForce-II/mmapeak by ReinForce-II on GitHub, posted by @jimtjames, 2026-10-11.
Expected: 150 to 200 TFLOPS on BF16 and FP16. About 83 TFLOPS on TF32 is normal; that is also what gpu-burn -tc reports. Source: @jimtjames, 2026-10-11 mmapeak loops matrix instructions on zero-filled data, so it measures peak issue rate and reads higher than a real GEMM. Source: repo github.com/ReinForce-II/mmapeak by ReinForce-II on GitHub, posted by @jimtjames, 2026-10-11. For comparison, the 170tune reference 8 GB card measured BF16 188 to 193, FP16 159 to 160 and TF32 89 to 92 TFLOPS in GEMM at stock. Source: repo github.com/cachenetics/170tune by cachenetics on GitHub, posted by @rh3dai, 2026-10-04.
torch-performance-test-data
A PyTorch square-GEMM benchmark with a public results table (Chinese README). Its table has one row labelled CMP170HX, contributed by GitHub user bendy2: FP32 12.15, TF32 77.09, FP16 166.48, BF16 177.91 TFLOPS (older v1 protocol, no host details). Two unlabelled "NVIDIA Graphics Device" rows nearby match the same profile but are not confirmed as 170HX. For scale, the same table lists an A100 SXM4 80GB at 258.74 FP16 and 264.53 BF16, and an RTX 3090 at 75.64 FP16. Source: repo github.com/zzc0721/torch-performance-test-data by zzc0721 on GitHub, posted by @jimtjames, 2026-10-11.
git clone https://github.com/zzc0721/torch-performance-test-data && cd torch-performance-test-data
uv sync && uv run test.py # or: python test.py
uv run test.py --device cuda:1 --sizes 2048 4096 8192
Needs torch 2.7 or newer and Python 3.10 or newer. Version 2 of the benchmark reports the median TFLOPS and its spread per matrix size; do not submit results with a spread (CV) above 5 percent. Source: repo github.com/zzc0721/torch-performance-test-data by zzc0721 on GitHub, posted by @jimtjames, 2026-10-11.
Caveats
- Neither tool is CMP-specific. mmapeak has no 170HX output in its repository and no licence file; its last commit is 2025-12-09. torch-performance-test-data has no licence file either. Source: repo github.com/ReinForce-II/mmapeak by ReinForce-II on GitHub, posted by @jimtjames, 2026-10-11.
Source: repo github.com/zzc0721/torch-performance-test-data by zzc0721 on GitHub, posted by @jimtjames, 2026-10-11. - Throughput says the tensor cores run at full rate, not that they compute correct results. Check correctness separately; see Testing and burn-in.