A Reconfigurable Approximate Computing RISC-V Platform for Fault-Tolerant Applications

Abstract

This paper presents "phoeniX," a novel reconfigurable embedded platform built upon the standard RISC-V ISA, designed to maximize energy efficiency through approximate computing while maintaining application-level accuracy. The platform uniquely allows the integration of diverse approximate circuits at the core level without requiring modifications to the core's control logic. Implemented as an optimized 3-stage pipelined RV32I(E)M architecture in 45nm CMOS, the core achieves remarkable metrics, including 1.89 DMIPS/MHz and an energy efficiency of 7.85 pJ per operation.

Report

Key Highlights

• Novel Platform: Introduction of "phoeniX," a highly reconfigurable embedded platform leveraging Approximate Computing (AC).
• RISC-V Standard: The platform utilizes the standard RISC-V Instruction Set Architecture (ISA).
• Decoupled AC Integration: Enables the seamless integration of approximate circuits with varying structures, accuracies, and timings directly into the core without requiring any changes to the core's underlying control logic.
• Configurable Trade-offs: Features novel control mechanisms allowing dynamic tuning of trade-offs between computational accuracy and energy consumption based on specific application needs.
• Target Applications: Demonstrated effectiveness through experiments involving image processing and the Dhrystone benchmark.

Technical Details

Implications

• Advancing Energy Efficiency: This research significantly pushes the boundary of energy-efficient embedded systems by providing a standardized, yet flexible, mechanism for incorporating approximate computing techniques directly into the RISC-V core.
• Modular Design for RISC-V: The ability to integrate AC circuits without altering control logic establishes a highly modular design methodology. This simplifies the development and adoption of heterogeneous and approximate cores within the RISC-V ecosystem.
• Enabling Fault Tolerance: By focusing on AC, which naturally tolerates minor errors for significant efficiency gains, the platform is ideally suited for critical fault-tolerant applications (e.g., edge AI, sensors) where high reliability and low power consumption are paramount.
• Competitive Performance: The reported metrics (1.89 DMIPS/MHz and extremely low energy consumption) position this architecture as a strong candidate for highly resource-constrained computing environments.