Chinese Startup Competes with Nvidia's AI Dominance Through New Chip Framework

At the recent Mobile World Congress in Barcelona, visitors observed a captivating Nvidia hologram at Dell Technologies’ booth. This event highlighted the growing interest in artificial intelligence (AI) technology and the competitive landscape surrounding it.

A new AI framework, named Chitu, has emerged from teams associated with China's Tsinghua University. This framework aims to lessen dependence on Nvidia chips for AI model inference, representing an important step for China toward greater technological self-sufficiency.

Innovative AI Framework

Chitu is designed as a high-performance inference framework specifically tailored for large language models (LLMs). It can function on chips that are manufactured in China, offering a viable challenge to the dominance of Nvidia's Hopper series graphics processing units (GPUs). This information was shared in a joint statement by the startup Qingcheng.AI and a research team led by Professor Zhai Jidong from Tsinghua University.

AI frameworks are vital as they provide a set of libraries and tools that enable developers to design, train, and validate advanced AI models efficiently. The newly released Chitu framework, which is now open-sourced, supports various mainstream models, including those from DeepSeek and Meta Platforms’ Llama series.

In practical tests involving the DeepSeek-R1 model using Nvidia's A800 GPUs, Chitu demonstrated impressive results. It achieved a remarkable 315 percent increase in model inference speed, while simultaneously reducing GPU usage by 50 percent as compared to existing foreign open-source frameworks.

Reducing Reliance on Foreign Technology

This initiative aligns with a broader strategy among Chinese AI firms to reduce their reliance on Nvidia. The US government has imposed export restrictions on Nvidia, prohibiting the sale of advanced H100 and H800 chips from the Hopper series to clients based in China, further motivating Chinese companies to seek alternatives.

DeepSeek, a rapidly growing AI company based in Hangzhou, has developed its models at significantly lower costs and with reduced computational resources compared to its Western counterparts. This rise has prompted discussions about a potential decline in demand for Nvidia GPUs within the Chinese market.

Founded in 2023, Qingcheng.AI is backed by the Beijing municipal fund dedicated to the AI industry, with Professor Zhai serving as the chief scientist. The startup has formed partnerships with leading Chinese GPU manufacturers such as Moore Threads, Enflame, and Iluvatar CoreX, according to CEO Tang Xiongchao's previous interviews with Chinese media.

Moreover, other technology companies in China have ramped up their efforts to lessen reliance on foreign technology in light of DeepSeek’s rapid advancements. In February, Infinigence AI, a computing infrastructure platform supported by Tsinghua talents and funding from major Chinese tech firms, announced a collaboration aimed at uniting the country’s top seven AI chip developers, including Biren Technology, Hygon Information Technology, and Huawei Technologies' Ascend.

Research efforts have also been fruitful. A recent study by scientists from ByteDance, the parent company of TikTok, reported a 170 percent increase in LLM training efficiency using an optimized system. This system has been implemented within some of ByteDance's production environments, resulting in significant savings of millions of GPU hours.

This development not only signifies a shift in the competitive landscape of AI technology but also showcases the determination of Chinese firms to innovate and build their capabilities in the face of external challenges.

AI, Startup, China