Deep Learning Toolkit-Driven Equivalence Checking of Flow-Based Computing Systems

Suraj Singireddy, Rickard Ewetz, Sumit Kumar Jha

Abstract: The processor-memory bottleneck inherent to von Neumann architectures has encouraged the development of alternative computing paradigms. One such paradigm is flow-based in-memory computing using nanoscale crossbars. Recent improvements to design automation tools has scaled the crossbar designs to over one million memristors and over one hundred input variables. However, the state-of-the-art verification method using graph reachability cannot verify the correctness of crossbar designs with more than twenty input variables. In this paper, we propose an equivalence checking technique that natively leverages existing deep learning infrastructure. We achieve this by observing an analogy between a memristor crossbar and recurrent neural networks (RNNs), which allows equivalence checking to be efficiently performed using neural network inference. Using benchmark circuits from the RevLib and MCNC suites, we show that our proposed method can, on average, verify the correctness of a design 166x faster than the state-of-the-art method.