[QST] Is there any fp16xfp16 GEMM sample using CUTE with a performance comparable to cublas?
Nobody has claimed this yet.
- Dominant language
- C++
- Stars
- 10.5k
- Forks
- 2.1k
- Avg merge
- 3d 11h
- Merged PRs (30d)
- 7
Description
What is your question?
I want to write my own fused fp16xfp16 gemm kernel with CUTE, but I can not find a tutorial/sample code with a performance comparable to cublas.
I noticed there are some tutorials in https://github.com/NVIDIA/cutlass/tree/main/examples/cute/tutorial, which has fp32xfp32 and int8xint8 gemm. But the performance of int8xint8 gemm is not good enough. I also noticed a 3rd party of fp16xfp16 gemm with CUTE https://github.com/leimao/CUDA-GEMM-Optimization?tab=readme-ov-file, but as shown in the readme, the performance is yet not comparable to cublas. So I wonder whether CUTE can give an official fp16xfp16 gemm kernel with good performance, so that I can develop based on that?
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by reviewing examples/cute/tutorial, especially the fp32xfp32 and int8xint8 GEMM samples, then compare the third-party CUDA-GEMM-Optimization results with cuBLAS. Done would be an official CUTE fp16xfp16 GEMM sample whose documented performance is comparable to cuBLAS.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp
- Domain
- performance
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 25/100