China's optical chip breakthrough boosts AI speed 100-fold using fraction of compute power

Chinese researchers have pioneered a novel all-optical interconnect system that could revolutionize artificial intelligence processing. Developed at Peking University, the system links standard electronic chips using light, dramatically accelerating AI inference speeds by over 100 times. Crucially, it achieves this performance leap while consuming only one-ninth of the computational power typically required.

The innovation addresses the insatiable demand for computational resources driven by the proliferation of AI models. Instead of relying on the conventional approach of adding more GPUs and expanding data centers, this new method offers a fundamentally different path. The research, published in the journal National Science Review, utilized Field-Programmable Gate Array (FPGA) chips, commonly employed in high-parallel-processing fields like missile guidance and autonomous driving.

“

As AI models permeate ever more applications, the industry’s appetite for computational power has grown insatiable.

The core of the system lies in custom-designed communication hardware components. These include a silicon photonic transceiver chip capable of operating at 400 gigabits per second, which efficiently converts electrical signals to optical and back. This optical linking bypasses many of the bottlenecks inherent in traditional electronic data transfer, enabling the significant speed and efficiency gains.

“

But a new study from Peking University suggests a radically different path: by optically linking chips with specific algorithms, they boost inference speeds by a factor of over 100 while slashing compute needs to just one-ninth.