20.3 A 23.9TOPS/W @ 0.8V, 130TOPS AI Accelerator with 16× Performance-Accelerable Pruning in 14nm Heterogeneous Embedded MPU for Real-Time Robot Applications

摘要

To solve social problems such as labor shortages, there are growing expectations for the advancement of human-cooperative robots. In such robots, advanced environmental recognition (mainly AI based), planning and control (normally non-AI algorithms) have to be processed simultaneously in real time. However, with current AI chips, meeting these objectives is challenging for several reasons: 1) A high-performance AI accelerator with 100TOPS peak class is required, in particular when a multi-camera system is employed for wide-range environment recognition. However, such an accelerator cannot be incorporated in a robot device because its power exceeds 10W [1] and a huge fan is required. 2) Embedded CPUs do not have sufficient performance for processing multiple robot tasks that mix AI and non-AI tasks in real-time. To solve these issues, we propose a power-efficient AI-MPU (microprocessor unit) including: 1) a flexible pruning rate control (named flexible N:M pruning) technology capable of up to 16$\times$ AI performance acceleration, and 2) a heterogenous architecture for multi-task & real-time robot operation based on the co-operation among a dynamically reconfigurable processor (DRP), AI accelerator (DRP-AI) and embedded CPU.