A team of scientists in China has unveiled the next generation of their fully optical artificial intelligence (AI) chip, named Taichi-II. This upgraded version builds upon the groundbreaking Taichi chip, first introduced in April 2024, and promises to push the boundaries of AI capabilities even further. Unlike conventional AI chips reliant on electronic components, Taichi-II harnesses the power of photons, or light particles, to perform its calculations. These photons activate tiny switches within the chip, enabling it to process information with exceptional speed and efficiency.
Taichi-II demonstrates a remarkable leap in performance compared to its predecessor. The new chip boasts a 40% improvement in accuracy for classification tasks, which involve sorting and identifying various types of data. Moreover, it achieves a staggering six orders of magnitude (a million-fold) increase in energy efficiency when operating in low-light conditions. This significant advancement is attributed to a novel training method called “fully forward mode,” where AI learning takes place directly on the optical chip without relying on digital simulations.
In this revolutionary approach, data flows in one direction only, unlike traditional training methods that involve iterative processing. The light passing through the chip interacts with tiny components, adjusting its direction, phase, and intensity, leading to immediate changes in the AI model’s parameters. This real-time learning process eliminates the need for repeated processing, contributing to the chip’s unparalleled speed.
Optical chips inherently possess several advantages over traditional chips. Their light-based nature makes them significantly less energy-intensive, allowing for much faster calculations. Photons, traveling at the speed of light, are able to process multiple signals simultaneously without generating heat, leading to faster and more efficient processing. Taichi-II, building upon these strengths, can be scaled much more effectively than competing designs, combining the benefits of various photonic technologies, including optical diffraction and interference.
The ultimate ambition of researchers is to use such photonically-powered AI chips to develop AGI models. These powerful systems would exhibit human-like intelligence and reasoning, capable of learning new skills beyond their initial training data. While this vision remains years away from realization, the modular architecture of Taichi offers a promising path forward. The chip’s design allows for the combination of multiple chiplets, effectively creating a powerful, scalable AI system.
In a demonstration, researchers successfully combined multiple Taichi chiplets, creating a system capable of simulating a network of nearly 14 million artificial neurons. This surpasses the performance of any other existing light-based chip design, which typically achieves around 1.47 million neurons. Furthermore, the combined system exhibits exceptional energy efficiency, performing over 160 trillion operations per watt of power consumed. This stands in stark contrast to earlier photonic chips, capable of managing only 3 trillion operations per watt, and conventional chips struggling to reach 10 trillion operations per watt.
Taichi-II marks a significant milestone in the development of optical AI chips, moving them from theoretical concepts to practical applications. The growing demand for high-power, low-energy computing solutions, crucial for developing advanced AGI models, makes this technology increasingly relevant. While the implications of such powerful AI systems raise concerns, Taichi-II represents a promising step towards a future where AI’s potential can be harnessed for the betterment of humanity.
