deepseek-ai/DeepEP
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepEP.git synced 2026-11-05 00:47:32 +00:00
Code Issues 308 Projects Packages Wiki Activity

[GH-ISSUE #633] Run the test_intranode.py test file, the bandwidth result is very large. #380

New issue
Open
opened 2026-10-26 09:59:47 +00:00 by bsakaguchi · 0 comments
bsakaguchi commented 2026-10-26 09:59:47 +00:00
Owner
Copy link

Originally created by @iMsuo on GitHub (May 13, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/633

Using two H100 GPUs (the machine has an 8-GPU setup), we tested test_intranode.py and increased the SM count（100）.
Image
The test bandwidth hit 500 GB/s, which exceeds the theoretical limit.

[tuning] SMs 100, NVL chunk 4: 357.23 GB/s (NVL), avg_t: 168.89 us
[tuning] SMs 100, NVL chunk 6: 403.38 GB/s (NVL), avg_t: 149.57 us
[tuning] SMs 100, NVL chunk 8: 431.41 GB/s (NVL), avg_t: 139.85 us
[tuning] SMs 100, NVL chunk 10: 440.36 GB/s (NVL), avg_t: 137.01 us
[tuning] SMs 100, NVL chunk 12: 461.04 GB/s (NVL), avg_t: 130.86 us
[tuning] SMs 100, NVL chunk 14: 447.61 GB/s (NVL), avg_t: 134.79 us
[tuning] SMs 100, NVL chunk 16: 433.24 GB/s (NVL), avg_t: 139.26 us
[tuning] SMs 100, NVL chunk 18: 442.79 GB/s (NVL), avg_t: 136.26 us
[tuning] SMs 100, NVL chunk 20: 432.81 GB/s (NVL), avg_t: 139.40 us
[tuning] SMs 100, NVL chunk 22: 431.02 GB/s (NVL), avg_t: 139.98 us
[tuning] SMs 100, NVL chunk 24: 430.64 GB/s (NVL), avg_t: 140.10 us
[tuning] SMs 100, NVL chunk 26: 410.76 GB/s (NVL), avg_t: 146.88 us
[tuning] SMs 100, NVL chunk 28: 411.30 GB/s (NVL), avg_t: 146.69 us
[tuning] SMs 100, NVL chunk 30: 411.45 GB/s (NVL), avg_t: 146.64 us
[tuning] SMs 100, NVL chunk 32: 407.09 GB/s (NVL), avg_t: 148.21 us
[tuning] SMs 100, NVL chunk default: 426.58 GB/s (NVL), avg_t: 141.44 us
[tuning] Best dispatch (FP8): SMs 100, NVL chunk 12, 461.04 GB/s (NVL), t: 130.86 us

[tuning] SMs 100, NVL chunk 4: 502.98 GB/s (NVL), avg_t: 232.63 us
[tuning] SMs 100, NVL chunk 6: 527.64 GB/s (NVL), avg_t: 221.76 us
[tuning] SMs 100, NVL chunk 8: 514.80 GB/s (NVL), avg_t: 227.29 us
[tuning] SMs 100, NVL chunk 10: 515.64 GB/s (NVL), avg_t: 226.92 us
[tuning] SMs 100, NVL chunk 12: 501.99 GB/s (NVL), avg_t: 233.09 us
[tuning] SMs 100, NVL chunk 14: 489.89 GB/s (NVL), avg_t: 238.85 us
[tuning] SMs 100, NVL chunk 16: 494.01 GB/s (NVL), avg_t: 236.86 us
[tuning] SMs 100, NVL chunk 18: 482.13 GB/s (NVL), avg_t: 242.70 us
[tuning] SMs 100, NVL chunk 20: 481.67 GB/s (NVL), avg_t: 242.93 us
[tuning] SMs 100, NVL chunk 22: 471.47 GB/s (NVL), avg_t: 248.18 us
[tuning] SMs 100, NVL chunk 24: 466.69 GB/s (NVL), avg_t: 250.73 us
[tuning] SMs 100, NVL chunk 26: 456.21 GB/s (NVL), avg_t: 256.48 us
[tuning] SMs 100, NVL chunk 28: 452.76 GB/s (NVL), avg_t: 258.44 us
[tuning] SMs 100, NVL chunk 30: 448.46 GB/s (NVL), avg_t: 260.92 us
[tuning] SMs 100, NVL chunk 32: 440.58 GB/s (NVL), avg_t: 265.58 us
[tuning] SMs 100, NVL chunk default: 466.06 GB/s (NVL), avg_t: 251.06 us
[tuning] Best dispatch (BF16): SMs 100, NVL chunk 6, 527.64 GB/s (NVL), t: 221.76 us

[tuning] SMs 100, NVL chunk 1: 174.69 GB/s (NVL), avg_t: 669.82 us
[tuning] SMs 100, NVL chunk 2: 309.17 GB/s (NVL), avg_t: 378.47 us
[tuning] SMs 100, NVL chunk 3: 395.33 GB/s (NVL), avg_t: 295.98 us
[tuning] SMs 100, NVL chunk 4: 445.00 GB/s (NVL), avg_t: 262.95 us
[tuning] SMs 100, NVL chunk 5: 468.89 GB/s (NVL), avg_t: 249.55 us
[tuning] SMs 100, NVL chunk 6: 491.77 GB/s (NVL), avg_t: 237.94 us
[tuning] SMs 100, NVL chunk 7: 503.19 GB/s (NVL), avg_t: 232.54 us
[tuning] SMs 100, NVL chunk 8: 495.00 GB/s (NVL), avg_t: 236.38 us
[tuning] SMs 100, NVL chunk 9: 490.27 GB/s (NVL), avg_t: 238.66 us
[tuning] SMs 100, NVL chunk 10: 489.83 GB/s (NVL), avg_t: 238.88 us
[tuning] SMs 100, NVL chunk 11: 483.32 GB/s (NVL), avg_t: 242.10 us
[tuning] SMs 100, NVL chunk 12: 484.54 GB/s (NVL), avg_t: 241.49 us
[tuning] SMs 100, NVL chunk 13: 480.66 GB/s (NVL), avg_t: 243.44 us
[tuning] SMs 100, NVL chunk 14: 470.20 GB/s (NVL), avg_t: 248.85 us
[tuning] SMs 100, NVL chunk 15: 475.32 GB/s (NVL), avg_t: 246.17 us
[tuning] SMs 100, NVL chunk 16: 466.02 GB/s (NVL), avg_t: 251.08 us
[tuning] SMs 100, NVL chunk default: 488.77 GB/s (NVL), avg_t: 239.40 us
[tuning] Best combine: SMs 100, NVL chunk 7: 503.19 GB/s (NVL), t: 232.54 us

What is the reason for this?

Originally created by @iMsuo on GitHub (May 13, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/633 Using two H100 GPUs (the machine has an 8-GPU setup), we tested test_intranode.py and increased the SM count（100）. <img width="1123" height="334" alt="Image" src="https://github.com/user-attachments/assets/b7f15ce3-6774-4f94-bc92-03152af33a93" /> The test bandwidth hit 500 GB/s, which exceeds the theoretical limit. [tuning] SMs 100, NVL chunk 4: 357.23 GB/s (NVL), avg_t: 168.89 us [tuning] SMs 100, NVL chunk 6: 403.38 GB/s (NVL), avg_t: 149.57 us [tuning] SMs 100, NVL chunk 8: 431.41 GB/s (NVL), avg_t: 139.85 us [tuning] SMs 100, NVL chunk 10: 440.36 GB/s (NVL), avg_t: 137.01 us [tuning] SMs 100, NVL chunk 12: 461.04 GB/s (NVL), avg_t: 130.86 us [tuning] SMs 100, NVL chunk 14: 447.61 GB/s (NVL), avg_t: 134.79 us [tuning] SMs 100, NVL chunk 16: 433.24 GB/s (NVL), avg_t: 139.26 us [tuning] SMs 100, NVL chunk 18: 442.79 GB/s (NVL), avg_t: 136.26 us [tuning] SMs 100, NVL chunk 20: 432.81 GB/s (NVL), avg_t: 139.40 us [tuning] SMs 100, NVL chunk 22: 431.02 GB/s (NVL), avg_t: 139.98 us [tuning] SMs 100, NVL chunk 24: 430.64 GB/s (NVL), avg_t: 140.10 us [tuning] SMs 100, NVL chunk 26: 410.76 GB/s (NVL), avg_t: 146.88 us [tuning] SMs 100, NVL chunk 28: 411.30 GB/s (NVL), avg_t: 146.69 us [tuning] SMs 100, NVL chunk 30: 411.45 GB/s (NVL), avg_t: 146.64 us [tuning] SMs 100, NVL chunk 32: 407.09 GB/s (NVL), avg_t: 148.21 us [tuning] SMs 100, NVL chunk default: 426.58 GB/s (NVL), avg_t: 141.44 us [tuning] Best dispatch (FP8): SMs 100, NVL chunk 12, 461.04 GB/s (NVL), t: 130.86 us [tuning] SMs 100, NVL chunk 4: 502.98 GB/s (NVL), avg_t: 232.63 us [tuning] SMs 100, NVL chunk 6: 527.64 GB/s (NVL), avg_t: 221.76 us [tuning] SMs 100, NVL chunk 8: 514.80 GB/s (NVL), avg_t: 227.29 us [tuning] SMs 100, NVL chunk 10: 515.64 GB/s (NVL), avg_t: 226.92 us [tuning] SMs 100, NVL chunk 12: 501.99 GB/s (NVL), avg_t: 233.09 us [tuning] SMs 100, NVL chunk 14: 489.89 GB/s (NVL), avg_t: 238.85 us [tuning] SMs 100, NVL chunk 16: 494.01 GB/s (NVL), avg_t: 236.86 us [tuning] SMs 100, NVL chunk 18: 482.13 GB/s (NVL), avg_t: 242.70 us [tuning] SMs 100, NVL chunk 20: 481.67 GB/s (NVL), avg_t: 242.93 us [tuning] SMs 100, NVL chunk 22: 471.47 GB/s (NVL), avg_t: 248.18 us [tuning] SMs 100, NVL chunk 24: 466.69 GB/s (NVL), avg_t: 250.73 us [tuning] SMs 100, NVL chunk 26: 456.21 GB/s (NVL), avg_t: 256.48 us [tuning] SMs 100, NVL chunk 28: 452.76 GB/s (NVL), avg_t: 258.44 us [tuning] SMs 100, NVL chunk 30: 448.46 GB/s (NVL), avg_t: 260.92 us [tuning] SMs 100, NVL chunk 32: 440.58 GB/s (NVL), avg_t: 265.58 us [tuning] SMs 100, NVL chunk default: 466.06 GB/s (NVL), avg_t: 251.06 us [tuning] Best dispatch (BF16): SMs 100, NVL chunk 6, 527.64 GB/s (NVL), t: 221.76 us [tuning] SMs 100, NVL chunk 1: 174.69 GB/s (NVL), avg_t: 669.82 us [tuning] SMs 100, NVL chunk 2: 309.17 GB/s (NVL), avg_t: 378.47 us [tuning] SMs 100, NVL chunk 3: 395.33 GB/s (NVL), avg_t: 295.98 us [tuning] SMs 100, NVL chunk 4: 445.00 GB/s (NVL), avg_t: 262.95 us [tuning] SMs 100, NVL chunk 5: 468.89 GB/s (NVL), avg_t: 249.55 us [tuning] SMs 100, NVL chunk 6: 491.77 GB/s (NVL), avg_t: 237.94 us [tuning] SMs 100, NVL chunk 7: 503.19 GB/s (NVL), avg_t: 232.54 us [tuning] SMs 100, NVL chunk 8: 495.00 GB/s (NVL), avg_t: 236.38 us [tuning] SMs 100, NVL chunk 9: 490.27 GB/s (NVL), avg_t: 238.66 us [tuning] SMs 100, NVL chunk 10: 489.83 GB/s (NVL), avg_t: 238.88 us [tuning] SMs 100, NVL chunk 11: 483.32 GB/s (NVL), avg_t: 242.10 us [tuning] SMs 100, NVL chunk 12: 484.54 GB/s (NVL), avg_t: 241.49 us [tuning] SMs 100, NVL chunk 13: 480.66 GB/s (NVL), avg_t: 243.44 us [tuning] SMs 100, NVL chunk 14: 470.20 GB/s (NVL), avg_t: 248.85 us [tuning] SMs 100, NVL chunk 15: 475.32 GB/s (NVL), avg_t: 246.17 us [tuning] SMs 100, NVL chunk 16: 466.02 GB/s (NVL), avg_t: 251.08 us [tuning] SMs 100, NVL chunk default: 488.77 GB/s (NVL), avg_t: 239.40 us [tuning] Best combine: SMs 100, NVL chunk 7: 503.19 GB/s (NVL), t: 232.54 us What is the reason for this?
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
hybrid-ep
chore/remove-clang-format-yapf
v2.5
nvDev
try-fix-cu128
mori-ep
antgroup-opt
compat-old-torch
add-linter
eager-rdma
v1.2.1
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/DeepEP#380
No description provided.