Skip to content
Book Open access

Adaptive Multiply-Accumulate Unit with Scalable Critical Path for Aggressive Voltage Underscaling DNN Accelerators

Jul 2026 · Design Automation Conference · 0 citations · 21 references
Computer Science

Abstract

Deep neural network (DNN) accelerators have widely been investigated for efficient inference, where matrix multiplication commonly implemented by multiply-accumulate (MAC) units dominates the power consumption. Underscaling the supply voltage of MACs can effectively reduce the power dissipation of DNN accelerators yet induces timing errors that may degrade the inference accuracy. In this paper, we propose a bit-width adjustment circuit (denoted as ADA) for arbitrary MAC units under aggressive voltage underscaling, which adaptively tunes the critical path as per input data patterns. Incurring 16.1% area overhead, the proposed ADA enables a MAC array to achieve zero accuracy loss under voltage underscaling ratios of 0.71 and 0.72 in the inferences of multi-layer perceptron (MLP) and convolutional neural network (CNN), respectively. While preserving accuracy, the MAC array equipped with ADA achieves up to 50.89% power reduction compared to that without ADA under nominal voltage operation. Furthermore, we propose ADA-Plus to specifically optimize the MACs in output stationary systolic arrays, which reduces the area of ADA by 23%. Compared with state-of-the-art error resilient circuit designs for voltage underscaling, to achieve a similar accuracy, the proposed adaptive MAC achieves the highest voltage underscaling ratio, resulting in 14.26%–48.44% power savings for the MAC array.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.