88Z AI, 30 September — DeepSeek has officially open-sourced infrastructure components for the Huawei Ascend platform, including TileLang, compute libraries, and distributed communication libraries, corresponding to NVIDIA platform components. TileLang supports DeepSeek V4 series model training, while other components cover core capabilities like matrix operations, with performance nearing hardware limits.
At a glance
- DeepSeek open-sources Ascend versions of TileLang and compute/communication libraries, corresponding to NVIDIA components
- Components support DeepSeek V4 training, with core performance in matrix operations nearing hardware limits
The update
DeepSeek has announced the official open-source release of an infrastructure component suite designed specifically for the Huawei Ascend computing platform. This suite includes the high-level language compilation tool TileLang, specialized compute libraries, and distributed communication libraries, with a functional architecture that directly corresponds to its previously released NVIDIA platform components. Notably, the Ascend version of TileLang has already been applied in the training of the DeepSeek V4 series models, successfully delivering high-performance operator support. Furthermore, key components such as DeepGEMM, DeepEP, and FlashMLA comprehensively cover core artificial intelligence computing capabilities, including matrix operations, distributed communication, and sparse attention mechanisms. Benchmark data indicates that these components operate with high efficiency, with performance levels approaching the theoretical limits of the relevant hardware, providing robust underlying technical support for the Ascend ecosystem.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- DeepSeekSource published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.