Impact of the Array Shape and Memory Bandwidth on the Execution Time of CNN Systolic Arrays

The use of Convolutional Neural Networks (CNN) has experienced a huge rise over the last recent years and its popularity has increased exponentially, mainly due to its application both for image recognition and certain applications related to artificial intelligence. The new applications of CNN request computing demands that are difficult to address by conventional processors.As a consequence, accelerators -both prototypes and commercial products- focusing on CNN computation have been proposed. Among these accelerators, those based on systolic arrays have acquired a special relevance; some examples are the Google's TPU and Eyeriss.Current research has focused on regular squared systolic arrays and most existing work assumes that there is enough memory bandwidth to feed the systolic array with input data. In this paper we explore the design of non-squared systolic arrays and address the impact of the memory bandwidth from a performance perspective.This work makes two main contributions. First, we found that some workloads with non-squared arrays achieve similar performance to systolic arrays twice as large, which can translate in area and/or energy benefits.Second, we present a performance comparison varying the main memory bandwidth for current DRAM devices. The analysis reveals that main memory bandwidth has a great impact on performance and that the decision of which technology use is key for the system performance. For the 64x64 array size it is necessary to use HBM2 memory to avoid the slowdown that would introduce cheaper technologies (e.g. DDR5 and DDR4).