New Delhi: Chinese AI company DeepSeek announced on Wednesday it had teamed up with Huawei Technologies to develop programming tools for the Ascend chips. The transfer is another indication of the growing cooperation among China’s tech giants to look for alternatives to Nvidia’s ecosystem.
DeepSeek has announced that it will open source all of the infrastructure for the Ascend platform, such as a high-level programming language called TileLang and the compute and communication libraries.DeepSeek has stated on its official WeChat blog that it will open source everything it has developed on the Ascend platform, including a high level programming language (TileLang), as well as compute and communication libraries.
A homegrown GPU software stack must begin with an easy-to-program language without compromising performance, the company said. Its words say that it is the first thing to have a high level language that would be universal, but would still be capable of getting the full potential out of the hardware. DeepSeek claims to have created TileLang with this in mind, and that it is simpler than Nvidia’s CUDA to program, making it more efficient and easier to simplify the logic of programming code.
Partnering is more than tuning the models. It focuses on the programming structures and methods which enable efficient execution of AI programs on various types of hardware. Huawei supported the rollout of the infrastructure, Investing.com reports, and together the two companies are advancing a new program called “supernode,” which is powered by 128 950 Ascend chips.
This is the latest in Huawei’s recent actions. It introduced a brand new family of chips for AI computing power just weeks ago, with plans to have its AI chips primarily used to train models by 2027. Huawei is very much a part of the Chinese government’s plan to build at home the chips it needs to compete with Nvidia.Huawei has been a key component of Beijing’s efforts to reduce dependence on Nvidia, with the company’s development of domestic AI chips.
Even for the Chinese leading AI lab, DeepSeek, they are still using a combination of Nvidia and Huawei chips. Thailand is to a large degree banned from selling Nvidia processors in its country and its V4-Flash variant was optimized for the Huawei processors.
One of Nvidia’s greatest assets has always been its software. The tools and code that developers have created to work with CUDA have been there for years and switching is expensive. An open-source language and library set for Ascend might help to drop that barrier, but it’s too soon to tell if developers will take to it.
Further releases were not disclosed by either company, nor was the financial terms for the future release.
