On the Accuracy of Analog Neural Network Inference Accelerators

Abstract

Specialized accelerators have recently garnered attention as a method to reduce the power consumption of neural network inference. A promising category of accelerators utilizes nonvolatile memory arrays to both store weights and perform in situ analog computation inside the array. While prior work has explored the design space of analog accelerators to optimize performance and energy efficiency, there is seldom a rigorous evaluation of the accuracy of these accelerators. This work shows how architectural design decisions, particularly in mapping neural network parameters to analog memory cells, influence inference accuracy. When evaluated using ResNet50 on ImageNet, the resilience of the system to analog non-idealities-cell programming errors, analog-to-digital converter resolution, and array parasitic resistances-all improve when analog quantities in the hardware are made proportional to the weights in the network. Moreover, contrary to the assumptions of prior work, nearly equivalent resilience to cell imprecision can be achieved by fully storing weights as analog quantities, rather than spreading weight bits across multiple devices, often referred to as bit slicing. By exploiting proportionality, analog system designers have the freedom to match the precision of the hardware to the needs of the algorithm, rather than attempting to guarantee the same level of precision in the intermediate results as an equivalent digital accelerator. This ultimately results in an analog accelerator that is more accurate, more robust to analog errors, and more energy-efficient.

Tunable Intervalence Charge Transfer in Ruthenium Prussian Blue Analog Enables Stable and Efficient Biocompatible Artificial Synapses

“…As most neural network weights are low significance (low-conductance), the hardware mapping for this scheme is natural and could lead to low overall error profiles at inference stage. [22] Meanwhile, the range clipping would not be a major detriment to learning performance, since a compressed range of 6-8 bits writable space is more than sufficient for most online learning applications using emerging non-volatile memory devices [23] even when considering write noise in the loop. [24] Endurance: The channel conductance can be repeatedly cycled using 10 6 switching pulses without altering switching characteristics (Section S5, Supporting Information).…”