RISC-V Vector Processors for Edge AI and High-Performance Computing: A Focused Review of Architectures and Performance Evaluation Methods

Abstract

Edge artificial intelligence and high-performance computing differ in available resources, numerical precision, and workload size, making the suitability of RISC-V vector processors dependent on their conditions of use. This paper presents a focused review of selected RISC-V vector processor designs and studies evaluating their performance, aiming to analyse architectural trade-offs and determine the comparability of performance and energy-efficiency results. The review analyses and compares the selected studies according to design characteristics, evaluated workloads, software environments, baselines, and measurement boundaries, distinguishing architectural studies from platform and application evaluations. Comparisons are classified as direct quantitative, conditional quantitative, or qualitative according to the agreement between the conditions underlying the specific results. The analysis shows that kernel acceleration does not necessarily translate into an equivalent improvement in the complete application, and that reduced precision and computational specialisation require performance gains to be interpreted alongside output quality and software support. Differences in energy-measurement boundaries and implementation types also prevent a unified ranking of all reviewed designs. The review contributes by linking results to their conditions and clarifying the limits of their use in evaluating architectural choices for both domains, rather than merely comparing reported peak values. The conclusions remain limited by the scope of the selected studies, their different configurations, and the relationship of some works to shared architectural families. This paper emphasises the importance of common workloads and baselines, documented operating conditions, and combined kernel and application benchmarks to improve the reliability of future comparisons.