DeepSeek open-sourced six Huawei Ascend projects whose matrix kernels hit 99.8% of the chip's peak speed.

this will help reduce reliance on Nvidia further for them.

The headline project is TileLang, the programming language DeepSeek writes its own GPU kernels in.

Most of the operators that trained its V4 models are written in TileLang, and each now has an Ascend version. That means the same source code can compile for Huawei chips instead of Nvidia ones.

a 2nd repo, DeepGEMM-Ascend, does matrix multiplication and reports 431 BF16 TFLOPS against a 432 ceiling.
DeepEP-Ascend, which moves data between chips, shipped with no license file while the rest are MIT, so nobody can legally reuse it yet.

Liang Wenfeng told investors Huawei could start delivering training chips this quarter.