Mapping Systolic Arrays onto 3D Circuit Structures: Accelerating Convolutional Neural Network Inference

摘要

In recent years, numerous designs have used systolic arrays to accelerate convolutional neural network (CNN) inference. In this work, we demonstrate that we can further speed up CNN inference and lower its power consumption by mapping systolic arrays onto 3D circuit structures as opposed to conventional 2D structures. Specifically, by operating in 3D space, a wide systolic array consisting of a number of subarrays can efficiently implement wide convolutional layers prevalent in state of the art CNNs. Additionally, by accumulating intermediate results along the third dimension, systolic arrays can process partitioned data channels in parallel with reduced data skew for lowered inference latency. We present a building block design using through-silicon vias (TSVs) for the 3D realization of systolic subarrays. We validate the 3D scheme using a 2.5D FPGA design and demonstrate that when mapped onto 3D structures wide systolic arrays can scale up in size without increasing wiring length in interconnecting subarrays. Further, by taking full advantage of 3D structures, we are able to pipeline inference across multiple layers of a CNN over a series of systolic arrays, dramatically reducing the inference time per input sample. These improvements lead to significantly reduced inference latency, which is especially important for real-time applications where it is common to process samples one at a time.