TRAM: Training Approximate Multiplier Structures for Low-Power AI Accelerators

TL;DR

TRAM introduces a joint optimization method for approximate multipliers and AI models, significantly reducing power consumption in AI accelerators while maintaining accuracy.

Contribution

It presents a novel approach that jointly optimizes multiplier structures and AI model parameters, unlike prior methods that treat them separately.

Findings

Achieves up to 25.05% power reduction in CNNs on CIFAR-10.

Reduces power by up to 27.09% in vision transformers on ImageNet.

Maintains small accuracy loss despite power savings.

Abstract

Reducing power consumption in AI accelerators is increasingly important. Approximate computing can reduce power consumption while keeping the accuracy loss small. Since multipliers are power-hungry components in AI models, this paper focuses on synthesizing low-power approximate multipliers (AxMs). Unlike prior works that design AxMs separately from AI model training, we present TRAM, which jointly optimizes the AxM structure and AI model parameters to lower power with small accuracy loss. Experiments show that compared to state-of-the-art AxMs, TRAM achieves up to 25.05% AxM power reduction on CNNs with CIFAR-10, and reduces power by up to 27.09% on vision transformers with ImageNet.