Research: Design and Evaluation of an Energy-Efficient CNN Inference Accelerator for Ed...

Research question: How can a custom FPGA-based accelerator architecture be designed to achieve energy-efficient CNN inference for edge-AI applications compared to existing state-of-the-art implementations?

Background & Motivation: Convolutional Neural Networks (CNNs) have become foundational in computer vision and edge-AI applications, but their computational complexity poses challenges for real-time, low-power inference on resource-constrained edge devices. FPGAs offer a promising platform for custom hardware acceleration due to their reconfigurability and potential for energy optimization.

Research Gap: While prior work has demonstrated FPGA-based CNN accelerators, there is limited systematic evaluation of architectural and algorithmic co-design strategies to maximize energy efficiency specifically for modern edge-AI workloads. Furthermore, quantitative comparisons of energy, performance, and accuracy trade-offs across design choices are sparse in open literature.

Approach & Expected Contribution: This project will investigate architectural optimizations (e.g., quantization, data reuse, memory hierarchies) for CNN inference on FPGA, proposing a novel accelerator design. The methodology involves literature review, architectural modeling, implementation using high-level synthesis (HLS), and benchmarking on standard CNN models (e.g., CIFAR-10, ImageNet subsets). Power and performance metrics will be collected using on-board instrumentation, and results compared to reference implementations. The expected contribution is a quantitative, reproducible evaluation framework and a prototype achieving improved energy efficiency for targeted edge scenarios.

Why it Matters: Improving the energy efficiency of AI inference at the edge is critical for pervasive autonomous systems, IoT, and mobile healthcare. This research will inform future edge-AI hardware design by identifying effective architectural strategies, with potential impact on battery life, environmental footprint, and deployment of intelligent systems in power-sensitive domains.