DeepSeek Announces a Stronger Software Integration with Huawei GPUs

DeepSeek has open-sourced six software components optimised for Huawei’s Ascend AI accelerators, extending its work with the Chinese chipmaker from adapting its models to building the underlying software stack needed to run them. 

The release covers programming, computation, communication, attention and data-processing components, and is intended to provide Ascend with counterparts to DeepSeek’s existing open-source infrastructure for NVIDIA GPUs.

The six components are TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect. 

TileLang is a high-level programming language and compiler for writing AI kernels, while DeepGEMM handles general matrix operations and DeepEP provides distributed communication for large-scale model workloads. 

TileKernels covers vector computation and memory access, while FlashMLA provides kernels for DeepSeek’s attention architecture, and DeepSelect handles data filtering.

DeepSeek said in a WeChat announcement that the Ascend versions correspond one-to-one with components it previously released for NVIDIA hardware. 

The Ascend implementation of TileLang wraps Huawei’s lower-level Ascend C instructions, allowing developers to use a higher-level programming model while retaining access to hardware-specific capabilities. 

DeepSeek said every TileLang operator used in its own model training now has a corresponding high-performance implementation on Ascend.

The company also said Huawei supported development and that the two companies jointly advanced a 128-chip supernode based on Ascend 950 processors, with optimisation work spanning both computation and communication. 

DeepSeek described TileLang as a simpler programming model than NVIDIA’s CUDA and said a high-level language that can reach hardware performance is a prerequisite for an independent AI-chip software ecosystem.

The release follows DeepSeek’s shift towards Huawei hardware with the launch of DeepSeek-V4 in April. Huawei said its Ascend 950-based supernode would fully support V4, while DeepSeek specifically adapted its model for Ascend systems. 

Reuters subsequently reported that demand for Ascend 950 chips increased among Chinese technology companies following V4’s release.

These developments come as US export controls continue to restrict Chinese access to advanced NVIDIA AI chips, with some products subject to case-by-case licensing and strict conditions. 

This has increased pressure on Chinese technology companies to develop domestic alternatives in both AI hardware and software.