For some time, application-specific integrated circuits (ASICs) have been used to efficiently accelerate computational performance in many areas of technology, including Internet of Things (IoT) devices. ASICs, however, are tailored to specific applications and lack flexibility. High-end field-programmable gate arrays (FPGAs) can achieve similar computational performance to ASICs with far more flexibility, and are increasingly being used for edge computing and IoT applications--some examples being unmanned aerial vehicles (UAVs) and autonomous vehicles with real-time image processing.
The authors propose a new architecture template based on convolutional neural networks (CNNs) targeting more economical but resource-constrained low-end FPGAs, which are far more likely to be suitable for IoT devices. These target devices need high computational performance, but also high energy efficiency--best bang for the buck, if you like. The authors describe a hardware template architecture for deep neural inference, which they call FeatherNet, for these low-end, resource-efficient FPGA platforms.
Background and related work in deep CNNs are discussed. The authors’ minimalist CNN system architecture and its operation are then described in detail and illustrated with relevant diagrams. The design methodology is covered in good detail, including aspects of the CNN model, design constraints, and optimization schemes. Finally, the authors present evaluation results for their proposed architecture, including the experimental setup, the evaluation metrics used, and their conclusions. A thorough list of references is also provided.
This is an interesting and timely piece of research in an area of growing interest, particularly with respect to real-time processing by small autonomous IoT devices where energy efficiency is paramount. The paper is well written and quite readable, even for a nonspecialist.