[GH-ISSUE #132] Reduce SM usage #116

New issue

Closed

opened 2026-10-26 09:54:34 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 09:54:34 +00:00

Owner

Originally created by @zhou9402 on GitHub (Apr 24, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/132

Hi, from my understanding, most of SMs are used to move data, one purpose is message aggregation. But as NIC/NVLINK can support MOS(million operations per second) now, is it possible, in normal mode, send data at per token granularity ? such that only lesser SM needed to issue transactions. And cooperatively, in NIC, use put_nbi_thread for each token instead pf put_nbi_warp to increase concurrency.

Originally created by @zhou9402 on GitHub (Apr 24, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/132 Hi, from my understanding, most of SMs are used to move data, one purpose is message aggregation. But as NIC/NVLINK can support MOS(million operations per second) now, is it possible, in normal mode, send data at per token granularity ? such that only lesser SM needed to issue transactions. And cooperatively, in NIC, use `put_nbi_thread` for each token instead pf `put_nbi_warp` to increase concurrency.

bsakaguchi closed this issue

2026-10-26 09:54:34 +00:00

bsakaguchi referenced this issue

2026-10-26 10:00:44 +00:00

[PR #116] [MERGED] fix: not output result in some linux system #435