Dataflow에 따른 Systolic Array의 연산 성능 분석

Abstract

Today, Deep Neural Networks (DNNs) have been widely used for various applications. Because the DNNs require a large amount of computation, hardware accelerators are commonly used to speed up the inference processing. In the systolic array architecture, a common hardware structure for neural network accelerators, the type of dataflow defines how data is stored in a processing element (PE) and exchanged among adjacent PEs. The computing performance of the systolic array differs depending on the type of the dataflows. Therefore, data flow analysis is crucial to maximize the inference performance. In this work, the computing performance depending on the data flows is evaluated using an open-source systolic array simulator called SCALE-Sim, Experimental results show that the inference latency differs up to 3.2 times depending on the type of the dataflow.