Chinese research institutions have detailed experimental photonic artificial intelligence accelerators that reportedly deliver extreme performance gains over conventional GPUs, but only for carefully constrained workloads. According to the teams, these light-based chips show substantial improvements in both speed and energy efficiency when running narrowly defined generative tasks. They reported China’s light-based AI chips offer 100x faster speed than Nvidia GPUs at some tasks, with the advantage clearest in image synthesis, video generation, and vision-oriented inference where the computation can be tightly structured.
The reported performance gap is rooted in a fundamental architectural shift from electrons to photons. Nvidia GPUs such as the A100 are built on electronic circuits that move electrons through transistors to execute programmable instructions, which provides broad flexibility but also leads to high power draw, significant heat, and reliance on advanced semiconductor nodes. The Chinese photonic designs instead use light-based signal processing, allowing photons to serve as the information carrier and enabling massive parallelism through optical interference instead of sequential digital execution. The researchers emphasize that the results come from laboratory evaluations rather than commercial deployments, and that the benchmarks are tailored to showcase the strengths of optical analog computation.
One of the highlighted chips, ACCEL, comes from Tsinghua University and combines photonic components with analog electronic circuitry in a hybrid architecture. It reportedly operates on older semiconductor manufacturing processes while achieving extremely high theoretical throughput figures measured in petaflops, although these operations are tied to predefined analog functions rather than general-purpose code. ACCEL is positioned for tasks such as image recognition and vision processing, which depend on fixed mathematical transformations and predictable memory access. A second system, LightGen, developed by Shanghai Jiao Tong University in collaboration with Tsinghua University, is described as an all-optical generative processor containing more than two million photonic neurons and supporting image generation, denoising, three-dimensional reconstruction, and style transfer. Experimental results reportedly show performance gains exceeding two orders of magnitude against leading electronic accelerators in terms of time and energy usage under constrained test conditions. The researchers note that these systems are not intended as drop-in GPU replacements for training large models or running arbitrary software, but as specialized analog machines for narrow computational domains, highlighting both the promise of optical computing and the sizable gap between lab prototypes and broadly usable artificial intelligence hardware.
