DeepSeek releases open-source tools for Huawei's Ascend platform

Technology company DeepSeek has announced the release of open-source infrastructure components tailored for Huawei's Ascend computing platform. The release includes the TileLang advanced language compiler, computation libraries, and a distributed communication library, mirroring components previously made available for Nvidia's platforms.

DeepSeek states that its TileLang programming language offers simplified coding and enhanced development efficiency compared to Nvidia's CUDA, allowing for fuller utilization of chip capabilities and hardware performance limits. The TileLang approach has been validated on Nvidia platforms and is currently used in the training of DeepSeek's V4 model series.

The newly released Ascend version of TileLang encapsulates Huawei's Ascend C low-level instructions, providing an advanced programming interface without sacrificing hardware performance. The company confirmed that every TileLang operator used in its model training now has a corresponding high-performance implementation on the Ascend platform.

The open-source release also features core computation and communication components for the Ascend platform, including DeepGEMM for accelerated matrix operations, DeepEP for efficient large-scale cross-device communication, and TileKernels for data processing. These components are designed for reusability across various scenarios, with performance metrics approaching hardware limits in key test cases.