Processing-in-SRAM Acceleration for Ultra-Low Power Visual 3D Perception
TimeTuesday, July 12th4:18pm - 4:42pm PDT
Location3000, Level 3
Event Type
Research Manuscript
Autonomous Systems (Automotive, Robotics, Drones)
Autonomous Systems
DescriptionReal-time ego-motion tracking and 3D structural estimation are the fundamental tasks in autonomous cyber-physical systems, and they can be realized by SOTA Edge-Based Visual Odometry (EBVO) algorithms. However, the intrinsic data-intensive process of EBVO faces the hurdle of memory-wall when deployed on low-end devices of compute-centric architectures. In this work, we attempt to leverage SRAM-based PIM to alleviate such a bottleneck, so as to optimize the EBVO system systematically from the algorithm layer, architecture and physical layer. The proposed multi-layer optimizations allow for high tracking accuracy, improve 16x processing speed and reduce 80x energy consumption compared to the CPU implementation.