Advanced Parallel Processing Technologies: 13th International Symposium, APPT 2019, Tianjin, China, August 15–16, 2019, Proceedings

Convolutional Neural Network (CNN) has gained significant attention in the field of machine learning, particularly due to its high accuracy in character recognition and image classification. Nevertheless, due to the computation-intensive and memory-intensive character of CNN, general-purpose processors which usually need to support various workloads are not efficient for CNN implementation. Therefore, a great deal of emerging CNN-specific hardware accelerators is able to improve efficiency. Although existing accelerators are significantly efficient, they are often inflexible or require complex controllers to handle calculations and data transfer. In this paper, we analyze classical CNN applications and design a domain-specific instruction set of 9 matrix instructions, called RV-CNN, based on the promising RISC-V architecture. By abstracting CNN into instructions, our design possesses a higher code density and provides sufficient flexibility and efficiency for CNN than general-purpose ISAs. Specifically, the proposed instructions are extended to RISC-V ISA as custom instructions. Besides, we also introduce micro-architectural optimizations to increase computational density and reduce the required memory bandwidth. Finally, we implement the architecture with the extended ISA and evaluate it with LeNet-5 on the datasets (MNIST, Caltech101, and Cifar-10). Results show that compared with the Intel Core i7 processor and Tesla k40c GPU, our design has 36.09x and 11.42x energy efficiency ratio and 6.70x and 1.25x code density respectively.