Key points
- DeepSeek open-sourced programming tools for Huawei chips
- The tools aim to reduce reliance on Nvidia's CUDA
- The two built a system on 128 Ascend chips
DeepSeek said it has open-sourced programming infrastructure for Huawei Technologies' Ascend chips, and described the goal as building "a new generation of independent, self-controlled GPU software ecosystems." The tools are meant to reduce developers' reliance on Nvidia (NVDA)'s CUDA, the software that ties much AI work to Nvidia hardware.
DeepSeek also said it worked with Huawei on a "supernode" system built on 128 Ascend 950 chips. DeepSeek posted the announcement to its official WeChat account on Wednesday, two weeks after Huawei unveiled its next generation of AI processors.
What DeepSeek released
The release includes compute and communication libraries and a high-level programming language called TileLang. Each component matches a tool DeepSeek had already released for Nvidia chips, so the two sets line up one to one. DeepSeek said Huawei provided full support in developing the infrastructure.
TileLang offers "a simpler programming model" than CUDA, DeepSeek said. "To build a new generation of independent, self-controlled GPU software ecosystems, the first priority is establishing a high-level language that is universal, easy to program, and still capable of reaching the hardware's full performance potential," the company said. "TileLang was created precisely to meet this need."
DeepSeek said it first tested the approach on Nvidia's platform and now uses TileLang for most of the computational operations involved in training its V4 models. The Ascend version wraps Huawei's low-level chip instructions, allowing developers to work in a higher-level language. DeepSeek says the approach can preserve performance.
What it means for Nvidia
CUDA is Nvidia's platform for programming its chips. Developers who have built software around it can face substantial work adapting that software to other hardware. DeepSeek says its tools are intended to reduce that dependence.
DeepSeek is one of China's most-watched AI developers, and it has been moving toward domestic hardware for months. The company earlier laid plans to install more than 160,000 Huawei Ascend chips at a data center in Inner Mongolia. Huawei has said it expects its AI systems to be widely used for model training next year.
Nvidia shares were up 1.0% at $229.45 in premarket trading, as of 9:25 a.m. ET on Wednesday. Nvidia's sales to China have already been cut by US export limits, and the company has said it expects little to no data-center revenue from the country.
