Improved Accuracy of Stochastic Accumulator Using Low-Cost Bitonic Sorter

This paper presents lightweight MAC hardware that transitions deep learning operations from binary to stochastic. Previous studies primarily conducted addition operations in the binary domain due to a significant accuracy drop in the stochastic adder. Consequently, a new calculation method called block-based stochastic computing (BSC) has been proposed, but it still shows a substantial accuracy drop. In contrast, the proposed stochastic computing-based MAC calculator achieves a 9% improvement in inference accuracy with 23% in area and 18% increase in power consumption compared to the BSC-MAC.