Energy-Efficient FPGA Framework for Non-Quantized Convolutional Neural Networks

2025-10-15

AI Summary

To address the challenge of efficiently deploying full-precision CNNs on FPGAs in embedded scenarios, this paper proposes a CPU-FPGA heterogeneous acceleration framework that supports Darknet models end-to-end without quantization. The framework restructures the computation pipeline, customizes the on-chip memory hierarchy, and introduces a co-scheduling mechanism to enable direct implementation of floating-point and fixed-point full-precision convolutions on FPGA. It achieves, for the first time, throughput (+12%) and energy efficiency (+9%) comparable to state-of-the-art quantized FPGA accelerators—without any accuracy loss—while outperforming general-purpose CPUs by 23×. By eliminating the conventional reliance on model quantization for FPGA-based CNN deployment, the framework establishes a new hardware acceleration paradigm for edge real-time AI, delivering both high accuracy and high energy efficiency.

Technology Category

• Machine Learning: Hardware-aware ML
• Computer Vision: Learning & Optimization for CV
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• User Modeling, Personalization and Recommendation: Federated recommendation systems and personalization

Abstract

The growing demand for real-time processing in artificial intelligence applications, particularly those involving Convolutional Neural Networks (CNNs), has highlighted the need for efficient computational solutions. Conventional processors, very often, fall short in balancing performance, power consumption, and latency, especially in embedded systems and edge computing platforms. Field-Programmable Gate Arrays (FPGAs) offer a promising alternative, combining high performance with energy efficiency and reconfigurability. The presented framework addresses the complex and demanding computations of CNNs on FPGAs maintaining full precision in all neural network parameters. Specifically, our framework is based on Darknet which is very widely used for the design of CNNs and allows the designer, by using a similar input to that given to Darknet, to efficiently implement a CNN in a heterogeneous system comprising of CPUs and FPGAs. When compared with the FPGA frameworks that support quantization, our solution aims to offer similar performance and/or energy efficiency without any degradation on the NN accuracy.

Problem

• Energy-efficient FPGA framework for non-quantized CNNs
• Maintains full precision in neural network parameters
• Achieves similar performance without accuracy degradation

Innovation

• FPGA framework for non-quantized neural networks
• Maintains full precision in all network parameters
• Achieves similar efficiency without accuracy degradation