A Novel Automate Python Edge-to-Edge: From Automated Generation on Cloud to User Application Deployment on Edge of...

Area of Science:

• Computer Engineering
• Artificial Intelligence
• Embedded Systems

Background:

• Deploying Deep Neural Networks (DNNs) on Internet of Things (IoT) edge devices demands specialized hardware and software expertise.
• Existing methods for edge AI deployment often present significant complexity for developers.
• System-on-Chips (SoCs) offer potential for on-device AI but require efficient DNN integration.

Purpose of the Study:

• To propose a novel, fully automated design framework for deploying DNNs on SoCs for IoT edge applications.
• To simplify the implementation of hardware-accelerated DNNs on Field-Programmable Gate Arrays (FPGAs).
• To provide an easy-to-use Python interface mimicking popular deep learning frameworks.

Main Methods:

• A three-phase methodology: customization (user specifies DNN layer optimizations), generation (cloud-based binary creation for FPGA and software), and deployment (edge SoC receives programming files and libraries).
• High-level Python interface abstracts hardware complexities.
• Leverages cloud infrastructure for generating FPGA and software binaries.

Main Results:

• An optimized DNN for the MNIST dataset achieved over 60x speedup compared to a software version on a ZYNQ 7020 SoC.
• The accelerated DNN consumed less than 0.43W, demonstrating significant power efficiency.
• The framework offers a superior trade-off between throughput, power consumption, and system cost compared to state-of-the-art solutions.

Conclusions:

• The proposed automated framework significantly simplifies DNN deployment for IoT edge applications on SoCs.
• It enables efficient hardware acceleration on FPGAs with reduced development effort.
• The methodology provides a compelling balance of performance, power efficiency, and cost for edge AI.