Quantization tolerant network design and performance estimation of computation-in-memory for energy-efficient 3D object detection inference

This work discusses network architecture, network training method, and a quantization method for achieving a low-energy and high-energy-efficient system, aiming at equipping robots and automobiles such as Autonomous Mobile Robot (AMR) with Computation-in-Memory (CiM) and performing 3D object detection with low energy consumption. Augmented Point Cloud VoxelNet (APCVN) is a network that improves inference accuracy by allowing a slight increase in computational complexity. Multi-Stage Quantization Aware Training (MSQAT) and U-Quantization (UQ) are a learning method and a quantization strategy, respectively, that improve the quantization tolerance of APCVN. Furthermore, quantitative calculations are conducted to estimate ADC energy consumption per inference, operations per second, energy efficiency, memory array area, memory capacity, and latency per inference in the assumed CiM system. Results show that by applying proposed methods, ADC energy consumption per inference is reduced by 8.7% and energy efficiency is improved by 1.6 times while maintaining high inference accuracy.