Abstract

BARVINN is an open-source DNN accelerator that enables deep learning inference at arbitrary precision using dedicated, bit-level configurable Processing Elements (PEs). The system is efficiently managed by a dedicated RISC-V CPU controller, achieving 8.2 TMACs of computational power on the Alveo U250 FPGA platform. This architecture provides critical run-time programmability, allowing seamless adjustment of quantization levels without requiring complex hardware reconfiguration.

Report

BARVINN: Arbitrary Precision DNN Accelerator

Key Highlights

• Arbitrary Precision: The primary innovation is the ability to perform DNN inference at arbitrary precision levels, supporting diverse quantization strategies.
• RISC-V Control: The accelerator is managed and controlled entirely by a RISC-V CPU, showcasing RISC-V's role in coordinating specialized high-performance hardware.
• Performance: The implementation achieves a combined computational power of 8.2 TMACs (Tera Multiply-Accumulate operations per second) on the Xilinx Alveo U250 FPGA.
• Run-Time Programmability: BARVINN enables changing quantization levels during execution without requiring time-consuming hardware reconfiguration, a significant advantage over many low-precision accelerators.
• Open Source: The entire project is available as open source, promoting community adoption and customization.

Technical Details

• Architecture: The accelerator comprises 8 dedicated Processing Elements (PEs).
• PE Configuration: PEs are specifically designed to be configurable at the bit level, enabling the arbitrary precision functionality.
• Deployment Platform: The accelerator was implemented and validated using the Alveo U250 FPGA platform.
• Software Toolchain: A specialized code generator tool processes Convolutional Neural Network (CNN) models provided in the standard ONNX format.
• Control Flow: The code generator produces an executable command stream that is fed directly to and interpreted by the RISC-V controller, orchestrating the PEs.

Implications

• RISC-V Ecosystem Expansion: BARVINN further solidifies RISC-V's position as a suitable, flexible, and open control core for complex heterogeneous compute systems, moving beyond simple embedded applications.
• Flexibility in AI Hardware: The arbitrary precision capability is crucial for the AI ecosystem, allowing developers to dynamically balance performance, power consumption, and model accuracy for different deployment scenarios (e.g., edge devices vs. data centers).
• Democratization of Accelerator Design: By being open source, BARVINN provides a reusable, high-performance reference design for other researchers and commercial entities looking to rapidly prototype specialized, customizable AI hardware based on RISC-V control.