Control Variate Approximation for DNN Accelerators

TL;DR

This paper presents a control variate technique to reduce error in low-precision DNN accelerators, enabling power savings with minimal accuracy loss without extensive retraining.

Contribution

It introduces a novel control variate method for variance reduction in approximate DNN accelerators, improving power efficiency while maintaining accuracy.

Findings

Achieves 24% power reduction in DNN accelerators.

Maintains same performance with only 0.16% accuracy loss.

Applicable to multiple DNN models on CIFAR datasets.

Abstract

In this work, we introduce a control variate approximation technique for low error approximate Deep Neural Network (DNN) accelerators. The control variate technique is used in Monte Carlo methods to achieve variance reduction. Our approach significantly decreases the induced error due to approximate multiplications in DNN inference, without requiring time-exhaustive retraining compared to state-of-the-art. Leveraging our control variate method, we use highly approximated multipliers to generate power-optimized DNN accelerators. Our experimental evaluation on six DNNs, for Cifar-10 and Cifar-100 datasets, demonstrates that, compared to the accurate design, our control variate approximation achieves same performance and 24% power reduction for a merely 0.16% accuracy loss.