Carbon nanotube analog tensor core accelerating edge computer vision

Abstract Computer vision requires intense tensor operations, primarily matrix multiplications, imposing substantial computation demands. Using hardware such as GPU and ASICs for acceleration offers a viable solution. However, their application at edge is constrained by complexity in the computing architecture and incompatibility with analog systems. Here, we prototype an analog tensor core based on carbon nanotube charge-trapping nonvolatile memory for edge computer vision acceleration. The memory, exhibiting ~100 linear, symmetric analog weights with fast programming (500 ns) and data processing (1 Mbit/s), enables compact tensor core design and facile programming operation for matrix multiplications of vectors and scalars in visual tasks. As demonstrations, the analog tensor core proves three-dimensional spatial transformation with an error of <2.8% and edge detection with a signal-to-noise ratio of >22 dB, underpinning the potential of our analog tensor core for performing computer vision tasks in, e.g autonomous driving, VR/AR, robot navigation, and industrial automation. Proof-of-concept analog tensor processing simulation in autonomous driving scenario using our analog tensor core in large scales expects to enable a processing speed of over 10,000 fps and offload over 10^7 times computation cost.