mirror of
https://github.com/deepseek-ai/DeepEP.git
synced 2026-11-05 00:47:32 +00:00
[GH-ISSUE #16] Question about NVlink max bandwidth #16
Labels
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
deepseek-ai/DeepEP#16
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @lishicheng1996 on GitHub (Feb 25, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/16
Hi, I'm comfused about the NVlink max bandwidth (160GB/s) in the readme.
But I see the NVlink 4.0 total bandwidth is 900 GB/s in other place.
And we run DeepEP intranode test on H20, the max bandwidth is 337.35 GB/s
@haswelliris commented on GitHub (Feb 25, 2025):
The H20 NVLink's bi-directional bandwidth is 900 GB/s, while the theoretical unidirectional bandwidth is only 450 GB/s. Considering the overhead introduced by the NVLink protocol and other factors, the achievable bandwidth may fall below 380 GB/s.
In your scenario, it may be essential to increase the number of SMs. For instance, you could set Buffer.set_num_sms(32) or choose a larger value to achieve enhanced bandwidth.
@sphish commented on GitHub (Feb 26, 2025):
H800 GPU servers are equipped with a reduced configuration of NVLink, featuring a nominal bandwidth of 200GB/s. Our benchmarks show that the actual achievable bandwidth peaks at approximately 160GB/s. In contrast, the H20 GPU servers you are using offer higher bandwidth capacity.