A Chinese startup claims to have created its own TPU to compete with NVIDIA. The only problem is that it is three years late

A Chinese startup called Zhonghao Xinying (known internationally as CL Tech) has come to the fore with a bold promise. The company claims to have developed an AI chip that not only circumvents Western intellectual property restrictions, but also outperforms NVIDIA’s A100 chip. Which is very good, but also a little bad.

Chana arrives. The chip in question has been named “Chana”, and according to SCMP we are dealing with a GPTPU (General Purpose Tensor Processing Unit). Unlike NVIDIA GPUs, aimed at accelerating AI workloads, this is an ASIC, that is, an application-specific integrated circuit designed from the ground up for neural network workloads.

promise. According to Zhonghao Xinying Chana, it offers up to 1.5 times the performance of the NVIDIA A100 based on the Ampere architecture. Not only that: it achieves that performance with 30% lower consumption. The startup highlights that the computational cost per unit would therefore be less than half of that offered by the A100 chips.

A little history of the company. Behind Zhonghao Xinying is Yanggong Yifan, an engineer formed at Stanford and the University of Michigan. He worked on the development of several generations of Google TPUs and also on the development of Oracle chips, and in 2018 founded this startup in Hangzhou together with Hanxun Zhengan engineer who worked at Samsung for several years. They were joined by other engineers from Microsoft, Oracle, NVIDIA, Amazon and Facebook, they indicate. on Baidu. We are therefore faced with several of those cases of “boomerang talent” with Chinese engineers who are forged in the US and then return to China to create solutions for their own industry.

Solutions that do not depend on the West. Yanggong affirms that its chip features “fully self-controlled IP cores, a custom instruction set, and a fully in-house computing platform. Our chips do not rely on foreign technology licenses, ensuring long-term security and sustainability from an architectural perspective.”

But. Although the achievement is striking, it is necessary to put it in perspective. The NVIDIA A100 is a 2020 AI GPU, and even with the improvements that this Chinese startup promises, its performance is, for example, far from H100 chips with Hopper architecture that appeared in 2022. Not to mention of the latest Blackwell Ultra chipswhich are currently NVIDIA’s greatest exponent in terms of AI chips. There are also no details about who makes the chip, and one of the candidates it would be SMICwhich has 7nm technology.

They are very far away, and they have another problem. The technical achievement of these engineers is certainly notable, but everything indicates that they are still far from what NVIDIA and its competitors are achieving. like AMD or Google with its recent TPU Ironwood. There is another element that works against them: Chinese manufacturers continue without having direct access to the most advanced photolithography on the market, and although it also there is progress from Chinese manufacturers in that sense, competing is certainly complicated without access to the most advanced technologies.

Pressure. In 2024 the company achievement revenues of 598 million yuan (73 million euros) with a net profit of 85.9 million yuan, but in the first half of the year the income was only 102 million yuan and had losses of 144 million yuan. The firm has reached an agreement with its investors by which it will have to go public at the end of 2026, or else it will be forced to buy back shares. The financial pressure is therefore notable for the company, which must demonstrate in the coming months that its roadmap is truly competitive.

In Xataka | China was no longer supposed to be able to get its hands on NVIDIA’s most advanced chips. Until he found a shortcut in Indonesia