DeepGEMM-Ascend: DeepSeek brings its matrix computation cores to Huawei chips, achieving 99.8 per cent of peak hardware performance

Released under the MIT licence, the library implements the DeepGEMM API for Ascend 950 NPUs in BF16, FP8 and FP4 — another step towards phasing out CUDA in China.

DeepSeek released DeepGEMM-Ascend on September 30, an adaptation of its DeepGEMM die multiplication library to Huawei's Ascend neural processors. The code, licensed by MIT, retains exactly the same interface as the version written for Nvidia GPUs: a developer can switch from one platform to another without rewriting their calls.

Near-ceiling performance

The library covers BF16, FP8 and FP4 formats, as well as specialized kernels (MQA, MegaMoE, GEMM attention logs, bundled). Developed and validated on the Ascend 950 series with the CANN 9.20 toolbox, it achieves up to 99.8% of the theoretical hardware usage on dense multiplications, and 1,701 TFLOPS in FP4 on a 4096×7168×16384 matrix.

This publication is part of a broader set of tools ported to Ascend by DeepSeek (including TileLang and DeepEP), at a time when Huawei is detailing its SuperPoD Flex architecture. The stated objective on the Chinese side is to reduce dependence on Nvidia's CUDA ecosystem.