Wire flash
DeepSeek open-sources six Huawei Ascend projects, matrix kernels hit 99.8% of theoretical peak
Editorial responsibility
- No named human review is recorded for this page.
- Source reporting is collected, normalized, translated or condensed automatically when needed.
- Automatically published source-backed update
DeepSeek has open-sourced six projects for Huawei's Ascend AI chips, achieving matrix kernel performance of 99.8% of the chip's theoretical peak speed. The headline project is TileLang, the programming language DeepSeek uses to write its own GPU kernels. Most operators that trained DeepSeek's V4 models are written in TileLang, and each now has an Ascend version, allowing the same source code to compile for Huawei chips instead of Nvidia ones. A second repository, DeepGEMM-Ascend, performs matrix multiplication and reports 431 BF16 TFLOPS against a theoretical ceiling of 432. DeepEP-Ascend, which handles inter-chip data movement, shipped without a license file while the rest use the MIT license, making it currently non-reusable. DeepSeek founder Liang Wenfeng told investors that Huawei could start delivering training chips this quarter, a move that could further reduce reliance on Nvidia.
Source report
DeepSeek has open-sourced six projects optimized for Huawei's Ascend chips, with matrix kernels reaching 99.8% of the chip's peak performance. This move is expected to further reduce the company's reliance on Nvidia hardware.
Key Projects
- TileLang – The headline project, TileLang is the programming language DeepSeek uses to write its own GPU kernels. Most operators that trained its V4 models are written in TileLang, and each now has an Ascend-compatible version. This allows the same source code to compile for Huawei chips instead of Nvidia ones.
- DeepGEMM-Ascend – A repository focused on matrix multiplication, reporting 431 BF16 TFLOPS against a theoretical ceiling of 432 TFLOPS.
- DeepEP-Ascend – Handles data movement between chips. Notably, this project was shipped without a license file, while the other five are under the MIT license. As a result, DeepEP-Ascend cannot yet be legally reused.
Licensing Note
All projects except DeepEP-Ascend are released under the MIT license. DeepEP-Ascend currently lacks a license file, making it non-reusable under standard open-source terms.
Industry Context
DeepSeek founder Liang Wenfeng told investors that Huawei could begin delivering training chips as early as this quarter.
Source
rohanpaul_aiNeutral / independent
Part of this Story
DeepSeek open-sources six Huawei Ascend projects, achieving 99.8% chip peak speed