FPGA-Based Accelerator for Quantized CNNs: High-Throughput Edge Deployment with Optimized Resource Utilization

This paper presents MaxNet, a high-throughput FPGA-based accelerator for quantized convolutional neural networks (CNNs), designed to meet the demand for efficient edge AI deployment on low-cost hardware. By targeting the Intel MAX 10 FPGA (10M08DAF484C8GES), MaxNet stands out from prior work focused on high-end platforms, offering a tailored solution for resource-constrained applications like IoT and embedded vision. The optimized two-layer CNN with 8-bit quantization (Q0.8 inputs/activations, Q1.7 weights) achieves 77% accuracy on CIFAR-10, with a throughput of 8,065 frames per second (0.124 ms/image) and power consumption of 1.2 W, using 861 LUTs (11%) and 9 M9K blocks (2.9%), as synthesized in Quartus Prime Standard Edition 22.1. Comprehensive power measurements, derived from Quartus Power Analyzer, psutil, and nvidia-smi, demonstrate superior efficiency compared to CPU (Intel Core i5-10400, 578 fps, 15 W) and GPU (NVIDIA GTX 1650, 730 fps, 50 W) baselines. Ablation studies validate the two-layer design and quantization choices, while sensitivity analysis optimizes clock frequency and numerical formats. Standardized tables and detailed figures clarify resource utilization and per-class accuracy, reinforcing MaxNet’s suitability for low-power, high-performance edge AI on cost-effective FPGAs.