Large AI chips from Nvidia and other manufacturers are nearing their physical scaling limits, according to Liao Heng, the chief scientist of Huawei’s semiconductor division. He expressed this view in a four-hour interview with a Chinese blogger, as reported by Bloomberg.
Liao stated that the industry has relied too heavily on increasing die size, transistor density, and high-bandwidth memory (HBM) capacity. HBM is used in AI accelerators for rapid data exchange between the processor and the computing system.
“There must be a limit to how they scale through larger computational dies and more HBM. The industry keeps pushing forward, but once it crosses this physical limit, an avalanche will occur,” the scientist said.
Huawei’s Alternative
Liao promoted Huawei’s approach called the Tau Scaling Law. This shifts the focus from continuously shrinking transistors to reducing signal transmission time between parts of the chip and the computing system.
Huawei introduced this approach in May at the Institute of Electrical and Electronics Engineers conference in Shanghai. At that time, He Tingbo, head of the company’s semiconductor business, stated that the technology should help Huawei achieve transistor density equivalent to 1.4 nm by 2031.
One element of the approach is LogicFolding, which the company describes as a method to distribute critical logic circuits across multiple layers to reduce signal transmission delays. According to Bloomberg, Huawei is preparing to unveil its first smartphone chip created using this technology.
Sanctions and Focus on Efficiency
Huawei is forced to seek alternatives due to U.S. restrictions. Since 2019, the company has been denied access to some Western technologies, and Chinese manufacturers cannot freely purchase advanced ASML lithography equipment for producing cutting-edge chips.
Liao linked the new approach not only to chip manufacturing but also to AI model architecture. He noted that Chinese companies have to compensate for the lack of computing resources with more complex design.
“We need to invest more effort in design to trade higher complexity for less computing resource consumption,” he said.
In this context, the scientist praised DeepSeek founder Liang Wenfeng. According to Liao, the company’s success is due to its need to find architectural solutions for training models with limited access to computing power.
Two Technological Ecosystems
Liao also stated that amid geopolitical divisions, each side will have to build its own production and technological chains.
“To survive, each side must create its own complete production and supply capabilities, even without serious confrontations between the two sides,” he said.
Bloomberg noted that Liao’s speech was another signal of Huawei’s confidence following the presentation of the Tau Scaling Law in May. The company is trying to show that U.S. restrictions not only forced it to seek alternatives but could also push the Chinese semiconductor industry toward a separate development trajectory.
In June, Huawei promised to release a new generation of Ascend AI chips every year and double their performance.
Earlier, Google DeepMind CEO Demis Hassabis suggested that China has nearly caught up with the U.S. and the West in the field of artificial intelligence.
