China develops experimental AI chip surpassing Nvidia A100 by 3000 times, but its application is limited

Chinese scientists have developed a chip that significantly outperforms existing high-performance artificial intelligence chips in tasks like image recognition and autonomous driving, as per a recent study. Although the new chip can't replace typical processors in computers or smartphones, plans are in motion to implement it in portable devices, electric cars, or smart factories soon. This move could enhance China's competitiveness in the widespread application of artificial intelligence, the researchers noted in their article.

The development comes as China aims to catch up with the USA in the AI race after Washington imposed several restrictions on the country's access to technologies, including advanced chips. The new chip, named ACCEL (All-Analogue Chip Combining Electronics and Light), relies on light and uses photons for computations and information transmission, enabling higher computational speeds.

The concept of a light chip isn't new, but existing chips still use electricity for computations due to the challenges in controlling photons. In lab tests, the new chip achieved a computational speed of 4.6 petaflops per second, which is 3000 times faster than one of the most common commercial AI chips, the Nvidia A100. Additionally, the Chinese chip consumes four million times less energy. However, this is just an experimental model, not a ready-for-mass-production processor.

The chip was developed by China's Semiconductor Manufacturing International Corporation (SMIC) using a cost-effective 20-year-old transistor manufacturing process. The performance can be improved by refining the manufacturing process or using more expensive sub-100-nanometer technological processes.

Unlike semiconductor chips, photonic chips utilize the physical properties of light, replacing transistors with ultra-microscopes and electric signals with light signals. Deploying photonic computing systems has been a complex task due to their intricate design and sensitivity to interference and systemic errors. The team introduced a computational architecture that combines photonic and analog electronic computations.

The use of light signals significantly enhances energy efficiency, and it's stated that the energy needed to run existing chips for an hour would power conventional chips for over 500 years. Low energy consumption could also help overcome the heat dissipation problem, a significant hurdle in further reducing the size of integrated circuits.

However, the analog computation architecture of the chip limits its application to specific tasks. It can't run various programs or compress files like typical chips with a broad range of functions. The tasks it can execute include:

• High-resolution image recognition

• Low-light computations

• Road traffic identification

So, powering generative AIs with these advanced processors is not possible yet.

The chip also has certain advantages in performing AI tasks for image processing as the ambient light itself carries information, allowing computations to be carried out directly during sensory perception.

Developing a new computational architecture for the AI era is a pinnacle achievement. However, a more critical task is applying this new architecture to practical applications, addressing major national and societal needs, which stands before us as a duty, said Dai Zonghai, one of the leaders of the research group.