Bespoke Co-processor for Energy-Efficient Health Monitoring on RISC-V-based Flexible Wearables

Abstract

This work introduces a highly energy-efficient, mechanically flexible RISC-V system designed for on-body health monitoring, addressing the challenges posed by the power constraints and limited gate counts of flexible electronics. The core innovation is the integration of a bespoke multiply-accumulate co-processor with fixed coefficients, optimized via a constrained programming approach for Multi-Layer Perceptron (MLP) inference. This bespoke design achieves a significant 2.35x speedup and 2.15x lower energy consumption compared to the state of the art, enabling near-real-time operation within existing flexible battery power budgets.

Report

Key Highlights

• Core Innovation: Integration of a bespoke, fixed-coefficient Multiply-Accumulate (MAC) co-processor into a mechanically flexible RISC-V architecture.
• Performance Metrics: Achieves an average 2.35x speedup and 2.15x lower energy consumption compared to existing flexible computing solutions.
• Footprint: The entire circuit occupies a highly compact area of only 2.42 mm².
• Application: Enables energy-efficient, near-real-time Multi-Layer Perceptron (MLP) inference for health monitoring directly on flexible wearables.
• Power Efficiency: The design operates strictly within the power budget limits of current flexible batteries.

Technical Details

• Base Architecture: A flexible RISC-V platform augmented with specialized hardware acceleration.
• Accelerator Design: A bespoke co-processor tailored for multiply-accumulate operations using fixed coefficients to maximize efficiency and compactness for specific ML models.
• Optimization Methodology: A constrained programming problem was formulated to simultaneously determine optimal co-processor constants and map MLP inference operations efficiently onto the custom hardware.
• Fabrication Context: The design specifically addresses hurdles common in flexible electronics, such as limited gate counts, large feature sizes, and high static power consumption.
• Goal: To create compact, model-specific hardware utilizing the low fabrication and non-recurring engineering (NRE) costs associated with flexible technologies.

Implications

• Advancing Flexible Wearables: This research provides a crucial step toward accessible, sustainable, and conformable healthcare devices capable of performing complex machine learning tasks locally, reducing reliance on power-hungry wireless transmission or external processing.
• RISC-V Customization: It powerfully demonstrates the intrinsic value of RISC-V—its extensibility. The ability to integrate a highly specialized, domain-specific accelerator (the bespoke MAC co-processor) directly alongside the core CPU shows how RISC-V enables tailored, energy-optimal solutions for constrained environments.
• Edge AI Expansion: By achieving near-real-time performance with extreme energy efficiency, the work pushes the boundaries of edge AI deployment, making sophisticated health monitoring classification practical on highly constrained, battery-powered devices.