China’s Moore Threads to unveil AI chip road map to rival Nvidia’s CUDA system

Moore Threads Technology is set to launch its latest graphics processing unit (GPU) architecture next week, fresh from the Chinese artificial intelligence chip designer’s stellar trading debut last Friday in Shanghai.\nFounder, chairman and CEO Zhang Jianzhong will unveil the Beijing-based firm’s fifth-generation GPU architecture during its inaugural MUSA Developer Conference on December 19 and 20 in the nation’s capital, according to the company’s latest post on WeChat.\nMUSA refers to the firm’s proprietary platform, Meta-computing Unified System Architecture – a Chinese-developed alternative to Nvidia’s widely used CUDA platform, for Compute Unified Device Architecture, released in 2007. These platforms provide the software layer that lets developers harness the power of GPUs for applications.\nIn his conference keynote, Zhang will also present Moore Threads’ MUSA road map, with a focus on full-stack development capabilities for enterprise users, according to the company’s post.\nThe event next week underscores Moore Threads’ efforts to step up its hardware and software capabilities, as well as grow its developer community, amid heightened competition with other Chinese GPU designers and Nvidia’s renewed push into the mainland market with its H200 chips after getting the green light from Washington.\n\nThe staging of the conference next week also reflects Moore Threads’ increased confidence after raising about 8 billion yuan (US$1.1 billion) in Shanghai, which ranked as the mainland’s second-largest initial public offering this year.\nMoore Threads’ shares soared 468 per cent in its market debut, as Chinese investors scrambled to own a piece of a company considered as the country’s best answer to Nvidia. Its shares rose 17 per cent to close at 735 yuan on Wednesday, reflecting investor excitement for the coming GPU architecture announcement.\nThe company’s previous generations of GPU architecture included Sudi, Chunxiao, Quyuan and Pinghu, which was introduced last year. With Pinghu, Moore Threads was able to support training and inference of AI models built with trillions of parameters.\nIts Pinghu-based GPUs have been used for pre-training of advanced large language models like those from DeepSeek, the chip designer had earlier said. Moore Threads was one of the initial members of the Model-Chips Ecosystem Innovation Alliance, which was formed in July to push the adoption of locally developed processors for AI projects.\nThe latest announcement by Moore Threads, which was established in 2020, appeared to bolster Zhang’s recent pledge to release a new-generation GPU every year to support China’s tech self-sufficiency and AI development initiatives.\n“Our goal is to become a leading GPU player with international competitiveness,” Zhang, Nvidia’s former China chief, told the official Shanghai Securities Journal in a report published last week.\n