China’s New AI Chip Unveils, Faster Than All

Just as the US government is imposing new restrictions on the sale of advanced GPUs to China, aiming to hinder its development in AI, researchers from Tsinghua University have achieved a surprising breakthrough.

These researchers have developed an all-analog photoelectronic chip that promises to revolutionize high-speed vision tasks. The chip, named ACCEL, combines electronic and light computing to achieve unprecedented energy efficiency and computing speed for vision-related processes.

Traditional digital computing units have long been limited by their energy consumption and computing speed when it comes to handling vision tasks. These tasks, such as image recognition for autonomous driving, robotics, medical diagnosis, and wearable devices, require high-resolution imaging, precise classification, and ultra-low latency.

The ACCEL chip takes advantage of the emerging field of photonic computing, which utilizes light to process information and perform computations. By integrating diffractive optical analog computing (OAC) and electronic analog computing (EAC) in a single chip, ACCEL achieves remarkable energy efficiency and computing speed.

The architecture of ACCEL.

Diffractive optical analog computing (OAC) is a method of using light and its properties to perform computations. It involves manipulating light waves through diffraction to encode and process information. By utilizing the interference patterns created by light, OAC can perform calculations in an analog manner, meaning it handles data continuously rather than in discrete digital steps.

On the other hand, electronic analog computing (EAC) is a computing approach that uses electronic components to perform calculations in a way that mimics continuous physical quantities. Instead of using digital signals represented by 0s and 1s, EAC operates with continuously varying signals. It leverages electronic circuits to process information in a manner similar to how physical systems, such as electrical circuits or natural phenomena, behave.

In a nutshell, OAC harnesses light waves to perform computations, while EAC uses electronic components to process information in a continuous manner. Both methods offer advantages for certain types of calculations and contribute to the development of high-speed vision tasks.

To illustrate the significance of ACCEL’s breakthrough, let’s consider the analogy of a camera. When we take a photo with a traditional digital camera, the captured image needs to be converted into digital signals using multiple analog-to-digital converters (ADCs). These ADCs consume considerable power and slow down the processing speed. ACCEL eliminates the need for these power-hungry converters by directly using light-induced photocurrents for calculations, resulting in a significantly reduced computing latency.

The capabilities of ACCEL are truly impressive. It achieves a systemic energy efficiency of 74.8 peta-operations per second per watt, which is more than three orders of magnitude higher than state-of-the-art GPUs or TPUs. Furthermore, its computing speed reaches 4.6 peta-operations per second, with over 99% of the computations performed optically.

ACCEL’s exceptional performance extends to practical applications as well. Through joint optimizations of optoelectronic computing and adaptive training, it achieves competitive classification accuracies in various tasks. For instance, it achieves accuracies of 85.5%, 82.0%, and 92.6% for Fashion-MNIST, 3-class ImageNet classification, and time-lapse video recognition tasks, respectively. Notably, ACCEL demonstrates superior robustness even in low-light conditions, making it suitable for wearable devices, autonomous driving, and industrial inspections.

Additionally, the operation of ACCEL at ultra-low power consumption will greatly alleviate heat generation issues, paving the way for comprehensive breakthroughs in future chip designs. Moreover, it provides a solid computational foundation for ultra-high-speed physical observations. This advancement brings significant benefits to demanding scenarios like unmanned systems and autonomous driving, where long-lasting endurance is crucial.

ACCEL can be used for ultra-low-power face wake-up schematic animation for electronic devices (Source: Tsinghua University)

The architecture of ACCEL is a remarkable feat of engineering. Unlike traditional optoelectronic computing systems that heavily rely on digital units, ACCEL combines diffractive optical computing and electronic analog computing in a scalable and flexible manner. Its architecture allows for scalability, nonlinearity, and adaptability, making it a practical solution for next-generation intelligent computing.

The development of the ACCEL chip marks a significant milestone in high-speed vision tasks. By harnessing the power of both photons and electrons, ACCEL offers unparalleled energy efficiency, computing speed, and system robustness.

This breakthrough by Tsinghua team holds profound significance not only for the practical application of optical computing technology but also for the integration of other high-performance computing techniques with existing electronic information systems.

Qionghai Dai, one of the corresponding authors of this research published in the prestigious journal Nature, emphasized, “Developing a computing system based on a completely new principle is a formidable challenge. However, it is even more crucial to successfully translate this next-generation computing architecture into real-world applications, addressing the critical needs of our nation and society.”

A special review of the research, published in Nature’s Research Briefing, further noted, “ACCEL might enable these architectures to play a part in our daily life much sooner than expected. “