Adaptive Block Floating-Point for Analog Deep Learning Hardware

Abstract

Analog mixed-signal (AMS) devices promise faster, more energy-efficient deep neural network (DNN) inference than their digital counterparts. However, recent studies show that DNNs on AMS devices with fixed-point numbers can incur an accuracy penalty because of precision loss. To mitigate this penalty, we present a novel AMS-compatible adaptive block floating-point (ABFP) number representation. We also introduce amplification (or gain) as a method for increasing the accuracy of the number representation without increasing the bit precision of the output. We evaluate the effectiveness of ABFP on the DNNs in the MLPerf™ datacenter inference benchmark-realizing less than 1% loss in accuracy compared to FLOAT32. We also propose a novel method of finetuning for AMS devices, Differential Noise Finetuning (DNF), which samples device noise to speed up finetuning compared to conventional Quantization-Aware Training.

An Electro-Photonic System for Accelerating Deep Neural Networks

“…In our work, we use the BFP number format, which provides a middle ground between FXP and FP number formats. BFP has been used for DNN inference [7], [13], [46] and training [17], [59] as it is less costly than the FP formats and achieves better accuracy than the FXP formats with the same bit-width. BFP format splits tensors into groups and assigns an exponent to each group that is shared by the elements within the group.…”

“…Combining BFP with analog hardware is a promising solution as BFP enables low-precision integer operations while providing a wider dynamic range than conventional integer arithmetic. This idea of using BFP formats for analog computing was first proposed by Basumallik et al [7] for DNN inference. In our work, we combine BFP and RNS to perform DNN training in photonic hardware.…”