DeepSeek released tools for Huawei's AI chips as an alternative to Nvidia

Chinese company DeepSeek has open-sourced a set of software tools for working with Huawei Ascend AI accelerators. The development is intended to simplify the creation and deployment of artificial intelligence models on Chinese processors and reduce the industry's dependence on Nvidia's technology ecosystem.

It concerns six software modules adapted for Huawei chips. In their purpose, they replicate the tools that DeepSeek previously used for Nvidia graphics processors.

One of the key components is TileLang, a tool for programming high-performance computing. Bloomberg calls it one of the Chinese tools meant to create an alternative to CUDA—Nvidia's software platform that has effectively become an industry standard for developing AI systems.

The software ecosystem remains one of Nvidia's main advantages. Even if a competitor can create a powerful processor, developers need libraries, compilers, and tools that allow neural networks to be efficiently trained and run on that hardware.

DeepSeek stated that it has made the tools for Huawei open and free to download. The company also reported working jointly with Huawei on optimizing software for the new Ascend 950 accelerators.

Huawei, for its part, has been actively developing its own CANN software platform for Ascend processors in recent years. In September, the company reported that the ecosystem already supports more than 90 large third-party open-source projects, including PyTorch, Triton, and vLLM.

This is important for the Chinese AI industry because of American restrictions on supplying the most powerful Nvidia processors. Chinese companies are increasingly using domestic accelerators, but for a mass transition they need not only to improve the specifications of the chips themselves, but also to create a compatible suite of software tools.

DeepSeek already uses Chinese processors in its own infrastructure. According to Bloomberg, the company has also worked directly with Huawei to improve the performance of its models on the Ascend platform.

Thus, China's competition with Nvidia is gradually moving from a battle solely over AI chip manufacturing to creating a full-fledged alternative ecosystem—from hardware to the software environment on which models run.

Based on materials from: Bloomberg, South China Morning Post, Huawei