China’s DeepSeek open-sources tools to help Huawei chips supplant Nvidia in AI

DeepSeek's release includes an Ascend-compatible version of TileLang, a custom programming language. TileLang is designed to optimize the development of high-performance kernels, which are crucial computation programs for GPUs and CPUs. This move expands TileLang's support beyond its primary Nvidia backend to officially include Huawei’s Ascend 950 accelerators.

The open-sourcing of these tools, particularly TileLang, addresses a critical aspect of chip independence. While hardware like Huawei's Ascend chips can be manufactured, their utility in advanced AI applications relies heavily on a robust and accessible software stack. Developing such a stack from scratch is a significant undertaking, requiring extensive engineering and developer adoption.

By providing tools that offer "native code generation, automatic scheduling, and synchronisation" for Ascend chips, DeepSeek is attempting to lower the barrier for developers to build and optimize AI models on Huawei's hardware. This initiative could foster a developer community around Ascend, which is essential for the long-term viability and competitiveness of any chip architecture, especially one seeking to challenge an established leader like Nvidia in its software ecosystem.