Chinese artificial intelligence (AI) company DeepThink announced on September 30 that it has released open-source software for developing AI programs based on Huawei's AI chips.
This move is seen as an effort to establish a Chinese AI software ecosystem in response to NVIDIA's CUDA.
According to reports from Chinese media, DeepThink made the software tools for Huawei's AI accelerator, Ascend (昇騰, Sheng Teng), available for free download through its official WeChat account. This includes TileLang, an advanced programming language necessary for AI development.
Bloomberg described TileLang as a Chinese AI development tool that corresponds to NVIDIA's CUDA, which helps developers create AI programs by leveraging the performance of NVIDIA's AI chips. NVIDIA has successfully built its GPU-centric AI ecosystem based on CUDA.
DeepThink explained that to create a new AI semiconductor ecosystem, a user-friendly programming language that maximizes chip performance is essential, which is the background for developing TileLang.
They noted that TileLang simplifies programming compared to CUDA, reducing the time developers spend writing code and simplifying program structure. It is also designed to better utilize the characteristics of AI chips, allowing them to operate closer to their maximum performance.
Specifically, the newly released TileLang for Huawei's Ascend allows developers to create AI programs with relatively simple code without having to handle complex chip instructions, thereby maximizing the performance of Huawei's AI chips.
In addition to TileLang, DeepThink has also released open-source software necessary for efficiently training and executing AI models, including computation, communication, and data processing tools. DeepThink reported that these software tools have achieved performance levels close to the maximum capabilities of Huawei's AI chips in several key tests.
DeepThink stated that it received active support from Huawei's research team during the development of the Ascend platform software. The two companies are optimizing the computational performance and inter-chip data communication of the next-generation AI chip, Ascend 950, while also jointly developing a massive system that connects 128 AI chips based on the Ascend 950.
Looking ahead, DeepThink plans to continue its technological development with Huawei to build an open AI software ecosystem.
Bloomberg noted that China is working to establish an ecosystem that replaces American AI chips and software in response to U.S. semiconductor export controls, with increasing collaboration between DeepThink, which develops AI models to compete with OpenAI and Anthropic, and Huawei, which produces AI chips for data centers.
This could help China reduce its dependence on NVIDIA chips and expand its capacity to develop and train AI models independently.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.