Design and benchmarking of a low-cost RISC-V-based high-performance computing cluster for edge computing

The exponential growth of IoT devices has led to a increase in data generation, necessitating real-time processing for critical decision-making, all while ensuring energy efficiency and privacy. Edge computing addresses these challenges by reducing latency through localized processing, but many edge devices still face limitations in real-time performance, energy efficiency, and scalability. In this work, we propose an energy-efficient, low-power HPC cluster based on single-board computers and open hardware. This cluster scales HPC capabilities with low-power, open-source solutions, making it ideal for IoT applications such as predictive maintenance in smart factories and real-time patient monitoring in healthcare. Our 10-node cluster, each featuring a StarFive JH7110 64-bit Quad-core CPU with an integrated GPU, achieves significant performance improvements. The computation time for a 1000x1000 matrix is reduced from 10.39 seconds on a single node to 1.637 seconds on 10 nodes, resulting in a 63% speedup. We introduce a custom distributed logistic regression algorithm designed for RISC-V, allowing model training on a cluster with improved scalability and performance. Our results show significant speedup in training time, with up to 5.6 times faster convergence as the number of nodes increases to 10. This work demonstrates the potential of RISC-V-based systems for efficient and distributed AI model training, making them well suited for decentralized applications in fields like academia, healthcare and finance while facilitating role of RISC-V in HPC.