Unlocking Efficiency in BNNs: Global by Local Thresholding for Analog-Based HW Accelerators

Abstract

For accelerating Binarized Neural Networks (BNNs), analog computing-based crossbar accelerators, utilizing XNOR gates and additional interface circuits, have been proposed. Such accelerators demand a large amount of analog-to-digital converters (ADCs) and registers, resulting in expensive designs. To increase the inference efficiency, the state of the art divides the interface circuit into an Analog Path (AP), utilizing (cheap) analog comparators, and a Digital Path (DP), utilizing (expensive) ADCs and registers. During BNN execution, a certain path is selectively triggered. Ideally, as inference via AP is more efficient, it should be triggered as often as possible. However, we reveal that, unless the number of weights is very small, the AP is rarely triggered. To overcome this, we propose a novel BNN inference scheme, called Local Thresholding Approximation (LTA). It approximates the global thresholdings in BNNs by local thresholdings. This enables the use of the AP through most of the execution, which significantly increases the interface circuit efficiency. In our evaluations with two BNN architectures, using LTA reduces the area by 42x and 54x, the energy by 2.7x and 4.2x, and the latency by 3.8x and 1.15x, compared to the state-of-the-art crossbar-based BNN accelerators.