DeepSeek Unveils Development Tools for Huawei AI Chips... Seeks to Reduce Reliance on NVIDIA

[Financial News] Chinese AI company DeepSeek jointly developed a programming tool for Huawei's AI chip, Ascend, with Huawei and released it as open source on the 30th.
DeepSeek said through its official WeChat account that day that it had released as open source programming infrastructure for the Ascend platform, including computing and communications libraries. DeepSeek explained that Huawei provided full support throughout the development process and that the two companies jointly optimized the computing and communications performance of a "supernode" solution consisting of 128 Ascend 950 chips.
DeepSeek said the components unveiled this time "correspond one-to-one" with the open-source components previously released for NVIDIA platforms.
The key component is TileLang, an open-source programming language for AI chips. DeepSeek said, "Compared with CUDA, TileLang is easier to program with, which can improve development efficiency and simplify code logic."
The announcement came two weeks after Huawei unveiled its next-generation AI chip and supernode system. At the time, Huawei said it had moved up mass production of its next-generation chip, the Ascend 960DT, to the first quarter of next year. It also expected its AI systems to be widely used for model training next year.


This initiative seeks to fill the development-tool gap, a weakness of Chinese-made chips, as they take on CUDA, regarded as NVIDIA's greatest competitive advantage in software ecosystems. It is the latest example of Chinese technology companies strengthening cooperation in search of an alternative to NVIDIA's ecosystem.
DeepSeek explained that building a self-controlled GPU software ecosystem requires, above all, an advanced language that is general-purpose and easy to program while extracting maximum performance from hardware. It said TileLang was created for that purpose.
It added, "All TileLang operators currently used by DeepSeek for training have corresponding high-performance implementations on Ascend as well." This means DeepSeek has laid the groundwork for moving model training to Huawei chips as well.
Cooperation between DeepSeek and Huawei is deepening.
Starting with the DeepSeek-V3.2 model released last September, DeepSeek supported Ascend chips on the day of each release. When it launched DeepSeek V4 in April, Huawei announced support for V4 across its entire lineup of Ascend supernodes.
DeepSeek is reportedly planning to deploy at least 160,000 Ascend 950DT chips at a data center it is building in the Inner Mongolia Autonomous Region. DeepSeek is also developing its own inference chips. It uses Huawei chips only for inference and continues to use NVIDIA chips for model training.

[email protected] Lee Seok-woo, international affairs specialist Reporter