Student Research Abstract: Real-Time 3D Registration and Fusion with SRAM-based Analog In-Memory Computing

In-memory computing systems are designed to offload workloads without transmitting tasks between CPU or GPU, reducing energy consumption. 3D sensors generate events of interest, usually powered by harvested energy or prolonged battery lifetimes. This work designed a three-dimensional (3D) sensing system using analog SRAM-based CIMs macro for storage and data processing to identify the subjects. The proposed neural network training scheme embedded the uncertainty model to tackle ADC non-deterministic thermal noise error. The design leverages the multiplication-accumulation (MAC) operation with limited precision on SRAM CIM macro, accelerating computation such as graph convolution, attention mechanism, and k nearest high dimensional feature vectors in Euclidean distance for the proposed registration pipeline. The experiment shows a reduction in the model size by up to 27 times, reduces the energy consumption by at least 100 times, and shortens the inference time by 297 times.