Compiler-Guided Mixed-Precision Risc-V Accelerator for Energy-Efficient Edge Vision Applications
Year: 2026
Type: article
Abstract: The deployment of deep neural networks at the edge is hard because edge devices have limited computational power and power budgets.Power efficiency and computation efficiency are critical for edge vision applications due to limited computational resources and restricted power budgets. Traditional fixed precision accelerators may not provide an optimal trade-off on a variety of network layers between energy and accuracy. This paper presents a Mixed-Precision RISC-V Accelerator built in the compiler guided way, leveraging Static Precision Analysis, Compiler Assisted Instruction Generation, Custom RISC-V Instructions Extensions and Mixed-precision Execution Engine to dynamically select proper numerical precision when executing Convolutional Neural Network Inference. The proposed architecture has been deployed on a Xilinx Artix-7 FPGA and validated with MobileNetV2, ResNet18 and EfficientNet-Lite benchmark models for edge vision applications. Experimental results show that the power saving is 29.8% with energy per inference reduction of 35.6%, inference latency is 24.3% lower, and energy efficiency is 41.2% compared to conventional fixed INT8 accelerators with accuracy degradation of less than 0.4%, corresponding to 98.7% classification accuracy. By only using 68% LUT, 61% DSP and 54% BRAM resources it showed the accelerator was ideal for real-time, low power embedded vision systems.
Author: Rajan.C
Institution: K.S. Rangasamy College of Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7219967855
