DeepSeek has open-sourced infrastructure components built for Huawei’s Ascend computing platform, extending its existing open-source stack beyond Nvidia hardware. The release covers the TileLang high-level language compiler toolchain, compute libraries, and distributed communication libraries, with each part mapped to a corresponding component previously released for Nvidia platforms.
According to DeepSeek, TileLang is designed as a general-purpose high-level language that simplifies programming while still reaching hardware performance limits. The company said the approach had already been validated on Nvidia and now underpins most operator implementations used in training the DeepSeek V4 model series. For Ascend, the newly released version wraps low-level Ascend C instructions and offers a higher-level programming model without sacrificing hardware performance.
DeepSeek also released DeepGEMM, DeepEP, TileKernels, FlashMLA, and DeepSelect, which target matrix operations, large-scale cross-device communication, vector computation and memory access, long-context sparse attention, and data selection. The company said key tests showed compute and communication performance close to hardware limits, and added that Huawei’s team supported the development process, including joint work on a 128-card supernode scheme based on Ascend 950 with deep optimization for compute and communication.
DeepSeek has formally open-sourced infrastructure components for Huawei’s Ascend computing platform, according to ChainCatcher. The release includes the TileLang high-level language compiler toolchain, compute libraries, and distributed communication libraries, with each component paired against an earlier open-source release for Nvidia platforms.
TileLang extended from Nvidia to Ascend
DeepSeek said TileLang is meant to provide a general-purpose high-level language that is easier to program while still reaching the upper bound of hardware performance. Compared with CUDA, the company said TileLang can improve development efficiency, simplify programming logic, and make use of chip-specific features through its programming model.
The TileLang path was first validated on Nvidia, where it already supports most operator implementations used in training the DeepSeek V4 model series. In the newly released Ascend version, DeepSeek said it wraps low-level Ascend C instructions and provides a high-level programming method without losing hardware performance.
DeepSeek added that every TileLang operator used in its training stack now has a corresponding high-performance implementation on Ascend.
Additional components released at the same time
The company also open-sourced DeepGEMM, DeepEP, TileKernels, FlashMLA, and DeepSelect. These components are used for general matrix operations, large-scale cross-device communication, standard vector computation and memory access, long-context sparse attention, and data selection, respectively.
Performance claims and Huawei collaboration
DeepSeek said compute and communication performance in multiple key tests came close to hardware limits. It also said Huawei’s team supported the development process, and that both sides worked together on a 128-card supernode scheme based on Ascend 950, with deep optimization for compute and communication.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.