DeepSeek ports its GEMM kernels to Huawei Ascend

DeepSeek open sourced its Ascend acceleration stack, starting with DeepGEMM-Ascend, a port of its DeepGEMM matrix multiplication library to Huawei’s Ascend platform. The library keeps the same package name and the same APIs as DeepGEMM on NVIDIA hardware, so the same code and development workflow carries over. It supports BF16, FP8 and FP4 GEMM, MQA logits and MegaMoE, under the MIT license.

The release note in the repository is dated 2026.09.30 and states support for Ascend 950 devices, with kernels designed to approach peak hardware performance. DeepSeek credits Huawei’s asc-comm team.

The benchmark tables in the readme are measured on an Ascend 950DT with CANN 9.20 and a cold L2 cache, and they are the substance of the release. Dense GEMM reaches up to 99.8% of the hardware limit, with BF16 by BF16 at 431 of 432 TFLOPS, FP8 by FP8 and FP8 by FP4 both at 99.5%, and FP4 by FP4 at 98.3%. The MQA logits kernel is described as bound by a fixed pipe rather than compute, saturating it at 99%. The MegaMoE kernel fuses expert parallel dispatch, two grouped GEMMs, SwiGLU and combine, reaching 846 TFLOPS at 384 experts.

Alongside it, DeepEP-Ascend provides expert parallel all to all for MoE dispatch and combine, with buffer APIs aligned to the NVIDIA DeepEP project and Ascend C kernels JIT compiled through DeepJIT. DeepSeek reports sustained dispatch at roughly 90% to 95% of physical payload bandwidth up to expert parallel degree 32.

What this is and is not, matters. It is infrastructure for running DeepSeek’s kernels on non-NVIDIA hardware, which is a real and useful contribution, and the API compatibility means it is genuinely low friction to adopt. It is not a new model, and it was measured on a proof of development kit. DeepSeek notes that Huawei’s Atlas 850E developer kit is planned for public availability around October 15, 2026 and is the recommended baseline for anyone reproducing these numbers. Huawei had already open sourced TileLang-Ascend in September 2025, so this is a broad port rather than a single breakthrough.

(Original source: DeepSeek on GitHub)