NVIDIA / NVIDIA/CUDALibrarySamples

LtMxfp8matmul tflops issue

Open
#283 0 comments 0 reactions 2 assignees View on GitHub

@rsdubtso is already working on this.

Since Sep 10, 2025.

cuBLASLt
Dominant language
Cuda
Stars
2.5k
Forks
478
PR merge metrics
No merged PRs in 30d

Description

Hi, I just use the example LtMxfp8Matmul in 5090, when I test 4096x4096x4096 matmul case, in only use 210+us, we can compute the tflops around 600+TFLOPs, but 5090 peak fp8 tlops is 419, is there something wrong?

the kernel name is:

cutlass3x_sm120_bstensorop_s16832gemm_block_scaled_ue8m0xe4m3_ue8m0xe4m3_f32_bf16_ue8m0xe4m3_128x128x128_1x1x1_0_tnn_align16_q_bias_bf16_relu_epiVs32n

I think the accumulate type should be fp32 right?....

Image

PS: I get this profile result by using Nsight system

Image

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.