A transprecision floating-point cluster for efficient near-sensor data analytics

Abstract

This paper presents a multi-core computing cluster based on the open-source RISC-V architecture, utilizing transprecision floating-point arithmetic for efficient near-sensor data analytics. The system achieves a minimum power budget through parallelization, sub-word vectorization, and near-threshold operation, while offering fine-grained precision tuning. This design demonstrates superior energy efficiency compared to the state-of-the-art, peaking at 162 Gflop/s/W using half-precision vectors.

Report

Key Highlights

• Transprecision Computing Focus: The primary innovation is a multi-core floating-point cluster leveraging transprecision computing principles for fine-grained precision tuning.
• Target Application: Designed specifically for efficient, low-power near-sensor data analytics, addressing the need for wide dynamic range and high precision.
• State-of-the-Art Energy Efficiency: The design significantly outperforms competitors, achieving peak energy efficiencies of 97 Gflop/s/W (single-precision scalars) and 162 Gflop/s/W (half-precision vectors).
• Comprehensive Support: Includes a full-fledged software stack, encompassing a parallel runtime and a dedicated compilation toolchain.

Technical Details

• Architecture Base: Built upon the open-source RISC-V architecture.
• Efficiency Techniques: System efficiency is maximized through three core techniques: parallelization, sub-word vectorization (SIMD-like operations), and near-threshold operation (running close to the minimum voltage for power saving).
• Precision Granularity: The cluster supports floating-point operations with tunable precision (transprecision), allowing users to adjust accuracy vs. power consumption.
• Validation: Design space exploration and experimental assessment were conducted using a cycle-accurate FPGA emulator, followed by post place-and-route power analysis.

Implications

• RISC-V Adoption in IoT/Edge: This work strongly validates the RISC-V ecosystem's capability to deliver highly energy-efficient and scalable solutions required for the exploding field of near-sensor computing and the Internet of Things (IoT).
• Shifting Arithmetic Paradigm: By successfully implementing transprecision computing, the paper suggests a viable path toward utilizing lower precision (like half-precision) for massive energy gains in real-world data analytics without sacrificing required accuracy.
• Open Hardware Advantage: Leveraging the open-source nature of RISC-V lowers development costs and facilitates future customization and incorporation of this highly efficient architecture into commercial or academic derivative projects.
• Software Enablement: The inclusion of a full software stack (runtime and toolchain) is crucial, enabling application developers to immediately utilize the hardware's complex parallelization and precision tuning features, thereby accelerating adoption.