deepseek-ai/DeepGEMM
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepGEMM.git synced 2026-11-05 00:47:26 +00:00
Code Issues 162 Projects Packages Wiki Activity

[GH-ISSUE #60] Weird benchmark results when m < 64 #37

New issue
Closed
opened 2026-10-26 21:11:02 +00:00 by bsakaguchi · 1 comment
bsakaguchi commented 2026-10-26 21:11:02 +00:00
Owner
Copy link

Originally created by @xiyanjoy on GitHub (Mar 13, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/60

m n k Time (us) TFLOPS GB/s
1 2112 7168 19.5 1.6 778.7
2 2112 7168 19.2 3.1 788.4
4 2112 7168 19.2 6.3 789.4
8 2112 7168 17.3 14.0 882.4
16 2112 7168 17.1 28.3 893.5
32 2112 7168 15.8 61.2 979.4
64 2112 7168 10.0 193.8 1586.6
128 2112 7168 11.5 337.1 1443.8
256 2112 7168 13.2 589.3 1372.6
1 7168 2048 11.1 2.6 1325.6
2 7168 2048 9.1 6.5 1618.6
4 7168 2048 8.4 14.1 1765.3
8 7168 2048 9.2 25.6 1612.5
16 7168 2048 8.7 53.8 1712.8
32 7168 2048 7.9 119.3 1931.0
64 7168 2048 7.2 260.9 2183.6
128 7168 2048 7.8 479.2 2139.4
256 7168 2048 10.3 732.4 1839.1

When m < 64, the DeepGemm will use 64 as block_m. So, in my opinion, the benchmark results of m<64 ought to be close to m=64 (the same access amount of weight, activation and scale, and the same computation flop). However, we observe worse behavior when benchmarking the cases of m<64.

Originally created by @xiyanjoy on GitHub (Mar 13, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/60 | m | n | k | Time (us) | TFLOPS | GB/s | |-------|-------|-------|-----------|--------|--------| | 1| 2112| 7168| 19.5| 1.6| 778.7| | 2| 2112| 7168| 19.2| 3.1| 788.4| | 4| 2112| 7168| 19.2| 6.3| 789.4| | 8| 2112| 7168| 17.3| 14.0| 882.4| | 16| 2112| 7168| 17.1| 28.3| 893.5| | 32| 2112| 7168| 15.8| 61.2| 979.4| | 64| 2112| 7168| 10.0| 193.8| 1586.6| | 128| 2112| 7168| 11.5| 337.1| 1443.8| | 256| 2112| 7168| 13.2| 589.3| 1372.6| | 1| 7168| 2048| 11.1| 2.6| 1325.6| | 2| 7168| 2048| 9.1| 6.5| 1618.6| | 4| 7168| 2048| 8.4| 14.1| 1765.3| | 8| 7168| 2048| 9.2| 25.6| 1612.5| | 16| 7168| 2048| 8.7| 53.8| 1712.8| | 32| 7168| 2048| 7.9| 119.3| 1931.0| | 64| 7168| 2048| 7.2| 260.9| 2183.6| | 128| 7168| 2048| 7.8| 479.2| 2139.4| | 256| 7168| 2048| 10.3| 732.4| 1839.1| When m < 64, the DeepGemm will use 64 as block_m. So, in my opinion, the benchmark results of m<64 ought to be close to m=64 (the same access amount of weight, activation and scale, and the same computation flop). However, we observe worse behavior when benchmarking the cases of m<64.
bsakaguchi commented 2026-10-26 21:11:03 +00:00
Author
Owner
Copy link

@soundOfDestiny commented on GitHub (Mar 13, 2025):

duplicate as https://github.com/deepseek-ai/DeepGEMM/issues/35

<!-- gh-comment-id:2720535368 --> @soundOfDestiny commented on GitHub (Mar 13, 2025): duplicate as https://github.com/deepseek-ai/DeepGEMM/issues/35
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
public-sync-20261014
0930-release
nv_dev
fix-mega-moe-fence
public-release-2609
public-release-2607
public-release-2606
mega-update
public-release-260416
ko3n1g/ci/build-wheels
nv_dev_f8e8fb5
nv_dev_950d31a
nv_dev_67fc648
nv_dev_c491439
nv_dev_4ff3f54
v2.1.1.post3
v2.1.1.post2
v2.1.1.post1
v2.1.1
v2.1.0
Labels
Clear labels   
pull-request
Mirrored from GitHub Pull Request

No labels
pull-request
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
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#37
No description provided.