Peking University: Neuromorphic Chip Beats Nvidia A100

Peking University researchers unveiled a neuromorphic chip that delivers up to 478 times faster performance than Nvidia A100 on real-time neural surface reconstruction by executing storage and computation inside the same memory array. The architecture removes the data-transfer bottleneck that forces traditional GPUs like the A100 to shuttle information between separate memory and processing units, directly addressing why AI training remains expensive and power-hungry. While the 478x gain applies only to this narrow workload and Nvidia H100 or B200 retain broad leads, the result on 40-nanometer silicon inside export limits shows frontier-level outcomes remain possible without cutting-edge hardware access, echoing DeepSeek’s January claim that better architecture can match frontier performance and trigger sharp market repricing.