An efficient open-source design and implementation framework for non-quantized CNNs on FPGAs

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 and graphical processing units (GPUs), 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. This paper presents a design and implementation framework for implementing CNNs seamlessly on FPGAs that maintains full precision in all neural network parameters thus addressing a niche, that of non-quantized NNs. The presented framework extends Darknet, which is very widely used for the design of CNNs, and allows the designer, by effectively using a Darknet NN description, to efficiently implement CNNs in a heterogeneous system comprising of CPUs and FPGAs. Our framework is evaluated on the implementation of a number of different CNNs and as part of a real world application utilizing UAVs; in all cases it outperforms the CPU and GPU systems in terms of performance and/or power consumption. When compared with the FPGA frameworks that support quantization, our solution offers similar performance and/or energy efficiency without any degradation on the NN accuracy. • An open-source library to accelerate CNN algorithms. • High design productivity, flexibility and adaptability of DarkNet design framework. • Achieve high performance by exploitation of the full parallelism of any FPGA. • Power efficiency of CNN inference on FPGAs for power-sensitive applications.