China’s light-based AI chips beat NVIDIA GPUs at some tasks by 100x

Chinese researchers have developed new photonic (light-based) AI chips, such as ACCEL and LightGen, that reportedly outperform NVIDIA’s GPUs by over 100 times in speed and energy efficiency for specific generative AI tasks like video production, image synthesis, and low-light vision. Unlike traditional NVIDIA GPUs, which use electrons flowing through transistors to execute flexible, general-purpose computations, these photonic chips perform preset analog mathematical operations via optical interference. This approach enables extremely fast and power-efficient processing but limits their flexibility and applicability to narrowly defined AI workloads. ACCEL, developed by Tsinghua University, is a hybrid chip combining photonic and analog electronic components, capable of delivering 4.6 PetaFLOPS while consuming minimal power. LightGen, created by a collaboration between Shanghai Jiao Tong University and Tsinghua University, is a fully optical chip with over 2 million photonic neurons, excelling in tasks like image generation, style transfer, and 3D image manipulation.