DeepSeek open-sources DeepEP-Ascend for Huawei Ascend NPUs
Start here
DeepSeek has open-sourced DeepEP-Ascend, a communication library for training and inference on Huawei Ascend NPUs. It provides expert-parallel all-to-all for MoE dispatch and combine, with FP8 dispatch and deferred epilogues, plus experimental pipeline parallelism, Bucket CP/DP collectives and Engram remote memory access. The public buffer APIs match the NVIDIA version of DeepEP. On an Ascend 950DT with CANN 9.2.0, DeepSeek reports dispatch at 373–375 GB/s at EP8 and 313–320 GB/s at EP128, while combine falls from 345–347 GB/s to 272–278 GB/s. The authors say larger EP sizes and combine are still being optimized. Those figures were measured on a proof-of-concept HDK supplied to DeepSeek with extra manual configuration. DeepSeek directs users to Huawei's Q3 commercial HDK release for Atlas 850E, planned for around October 15, 2026, via Huawei's Atlas 850E software download page, where access can be requested.
Generated from reports · updated 1 h ago
Report timeline
Follow the reports to see the event from different sides.
- Hacker News · AIDeepSeek releases DeepEP-Ascend, an expert-parallel communication library for Huawei Ascend NPUs
DeepSeek released DeepEP-Ascend, an open-source communication library for training and inference on Huawei Ascend NPUs, offering expert-parallel all-to-all for MoE dispatch and combine with FP8 dispatch and deferred epilogues, plus experimental PP, Bucket CP/DP collectives and Engram remote memory access. Measured on Ascend 950DT with CANN 9.2.0, dispatch reaches 373–375 GB/s at EP8 and 313–320 GB/s at EP128, while combine drops from 345–347 GB/s to 272–278 GB/s. The public buffer APIs align with the NVIDIA version of DeepEP, and the authors note that larger EP sizes and combine are still under optimization.
Heat for this event
Not enough continuous observations yet to draw a trend.