This proposed architecture introduces a novel matrix extension and customized quantization instructions for Risc-V CPUs, specifically for general convolutional neural network (CNN) applications. This presentation emphasizes the design of the matrix multiplication/accumulation instructions, aiming to achieve scalability/portability across diverse VLEN machines (VLEN agnostic). The key objectives include higher computing capacity, intensified compute intensity and reduced memory access bandwidth requirements. Furthermore, this work proposes an associated 2D-load/store unit (LSU) for matrix tiling enhancements and the Zero-Overhead Boundary handling to streamline user configuration cycles. Additionally, a novel quantization instruction is introduced, contributing to the acceleration of the entire CNN computations. By synergistically integrating these state-of-the-art techniques, the architecture demonstrates significant performance enhancements. Preliminary performance data underscores the benefits and the potential acceleration of General Matrix Multiply (GeMM) and convolutional neural network (CNN) workloads. Notable performance metrics include kernel loop MAC utilization rate surpassing 75% and compute intensity up to 9.6 (VLEN 512), achieved through advanced software unrolling techniques.

Details and slides: https://riscv-europe.org/summit/2024/conference#enhancing-convolutional-neural-network-computation-with-integrated-matrix-extension-details