Efficient, Accurate, and Flexible PIM Inference through Adaptable Low-Resolution Arithmetic

Abstract

Processing-In-Memory (PIM) accelerators have the potential to efficiently run Deep Neural Network (DNN) inference by reducing costly data movement and by using resistive RAM (ReRAM) for efficient analog compute. Unfortunately, overall PIM accelerator efficiency and throughput are limited by area/energy-intensive analog-to-digital converters (ADCs). Furthermore, existing accelerators that reduce ADC area/energy do so by changing DNN weights or by using low-resolution ADCs that reduce output fidelity. These approaches harm DNN accuracy and/or require costly DNN retraining to compensate.

To address these issues, this thesis explores tradeoffs around ADC area/energy and develops optimizations that can reduce ADC area/energy without retraining DNNs. We use these optimizations to develop a new PIM accelerator, RAELLA, which can adapt the architecture to each DNN. RAELLA lowers the resolution of computed analog values by encoding weights to produce near-zero analog values, adaptively slicing weights for each DNN layer, and dynamically slicing inputs through speculation and recovery. Low-resolution analog values allow RAELLA to both use efficient low-resolution ADCs and maintain accuracy without retraining, all while computing with fewer ADC converts.

Compared to other low-accuracy-loss PIM accelerators, RAELLA increases energy efficiency by up to 4.9x and throughput by up to 3.3x. Compared to PIM accelerators that cause accuracy loss and retrain DNNs to recover, RAELLA achieves similar efficiency and throughput without expensive DNN retraining.