[GH-ISSUE #197] Tested tests/test_fp8.py on SM100 with very poor performance #112

Closed
opened 2026-10-26 21:12:53 +00:00 by bsakaguchi · 0 comments
Owner

Originally created by @wasamtc on GitHub (Sep 24, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/197

Hello, I installed DeepGEMM using develop.sh and install.sh on an SM100 GPU, but when running tests/test_fp8.py, the performance was very poor with only single-digit FLOPS and very low bandwidth. However, I tested tests/test_fp8.py using the same steps on an SM90 GPU, and the performance was normal. Below are the results from the SM100

Image Below are the results from the SM90 Image my nvcc version Image my torch version Image
Originally created by @wasamtc on GitHub (Sep 24, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/197 Hello, I installed DeepGEMM using develop.sh and install.sh on an SM100 GPU, but when running tests/test_fp8.py, the performance was very poor with only single-digit FLOPS and very low bandwidth. However, I tested tests/test_fp8.py using the same steps on an SM90 GPU, and the performance was normal. Below are the results from the SM100 <img width="1940" height="827" alt="Image" src="https://github.com/user-attachments/assets/7165a8fa-aac6-46e6-8cc6-748dcd5553e7" /> Below are the results from the SM90 <img width="1951" height="541" alt="Image" src="https://github.com/user-attachments/assets/92ffbf20-a3d8-45c8-8d1a-4b13ed25b57a" /> my nvcc version <img width="848" height="189" alt="Image" src="https://github.com/user-attachments/assets/885ae239-73a3-4352-9532-24221fd7094c" /> my torch version <img width="533" height="85" alt="Image" src="https://github.com/user-attachments/assets/46c2ab4b-73dd-4453-9cb6-ce1e17ce1987" />
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepGEMM#112
No description provided.