Vicuna2.0: RISC-V Embedded Vector Unit with Half-Precision Floating-Point Support for TinyML

Abstract

With increasing computation demands on resource-constrained embedded systems, half-precision floating-point has emerged as a possible datatype to make use of the dynamic range of floating-point representations while reducing overall design complexity. The RISC-V vector extension includes a sub-extension for half-precision floating point, but its viability for embedded systems has not be evaluated. After implementing the Zvfh extension and all required dependencies, a design-space exploration was performed to find an optimal system for a common machine learning task, anomaly detection. When compared to an integer-only vector system, the half-precision floating-point system had a 25% lower LUT overhead due to the change in program characteristics and use of a smaller datatype for intermediate values with only a 8% higher execution time.