to leave a comment.
Chinese artificial intelligence (AI) company Deepseek announced on the 30th that it has co-developed and open-sourced programming tools for Huawei's AI chip 'Ascend' with Huawei.
This is an attempt to fill the void in development tools, which is a weakness of Chinese-made chips, in response to NVIDIA's CUDA software ecosystem, considered its greatest competitive advantage.
Reuters reported that this is the latest example showing Chinese technology companies are strengthening cooperation as they seek alternatives to the NVIDIA ecosystem.
Deepseek stated on its official WeChat account that it is open-sourcing programming infrastructure for the Ascend platform, including computation and communication libraries.
Deepseek explained that Huawei provided full support for the development process, and both companies jointly optimized the computation and communication performance of a 'supernode' solution composed of 128 Ascend 950 chips.
Deepseek explained that the components unveiled this time "correspond one-to-one" with the open-source components previously released for the NVIDIA platform.
The key is 'TileLang', an open-source programming language for AI chips. Deepseek stated, "Compared to CUDA, TileLang is simpler to program, can increase development efficiency, and simplify code logic."
Deepseek emphasized that to build an independently controllable Graphics Processing Unit (GPU) software ecosystem, a general-purpose, easy-to-program, high-level language that can maximize hardware performance is primarily needed, and TileLang was created for this purpose.
In particular, Deepseek stated, "Currently, all TileLang operators used by Deepseek for training have corresponding high-performance implementations on Ascend." This is interpreted to mean that the foundation has been laid to shift even model training to Huawei chips.
The cooperation between Deepseek and Huawei is deepening.
Deepseek supported Ascend chips from the day of release for its 'V3.2' model unveiled in September last year, and when 'V4' was launched in April, Huawei announced V4 support for its entire Ascend supernode product line.
Bloomberg reported earlier this month that Deepseek plans to deploy at least 160,000 Huawei 'Ascend 950DT' chips in the data center it is building in Inner Mongolia.
However, Bloomberg also reported that Deepseek uses Huawei chips only for inference, and still uses NVIDIA chips for model training. Deepseek is also reportedly developing its own inference chips.
This announcement came two weeks after Huawei unveiled its next-generation AI chips and supernode system. At the time of the unveiling, Huawei accelerated the mass production schedule for its next-generation chip 'Ascend 960DT' to the first quarter of next year, and expected its AI systems to be widely used for model training next year.
Newsletter
Get key news delivered to your email every morning
to leave a comment.