The computer chips use light to do AI tasks faster and with less energy. They could help make future computers work better and quicker.
Computer chips that combine light and electricity have been shown to boost computing performance while using less energy than traditional electronic chips. Two papers published highlight these photonic computing chips as a possible way to meet the growing demands of artificial intelligence.
As AI and deep learning models become more complex, traditional electronic computing is struggling to keep up—especially due to rising energy use. Photonic computing, which uses photons instead of electrons, could help overcome these limits.
Key operations in AI, such as multiplication and accumulation, can be done faster and more efficiently with photonic circuits. Until now, it has been hard to show this in real hardware combining photonics and conventional electronics. But the two new studies test the performance of such integrated chips.
In one study, the team presented a photonic accelerator called PACE, which enables extremely low-latency computing—important for real-time tasks.
A large-scale photonic accelerator, built with over 16,000 photonic components arranged in a 64 x 64 matrix, has achieved high-speed computing of up to 1 GHz. It also shows a significant reduction in minimal latency—up to 500 times lower—compared to smaller circuits or individual photonic components. This system is capable of solving complex computational tasks known as Ising problems, indicating its potential for practical use.
Separately, another photonic processor, made up of four 128 x 128 matrices, has been used to run artificial intelligence models with high accuracy. It successfully executes the natural language processing model BERT and the image-recognition neural network ResNet, achieving performance on par with traditional electronic processors.
This processor has also demonstrated the ability to generate text in the style of Shakespeare, classify movie reviews with precision, and play classic Atari games such as Pac-Man.
Both systems are considered scalable, although further improvements are still required. Together, they represent a step closer to using light-based systems for more powerful and energy-efficient computing.
