[Paper Review] Embedded Development Boards for Edge-AI: A Comprehensive Report

Hamza Ali Imran, Usama Mujahid+3 more|arXiv (Cornell University)|Sep 2, 2020

IoT and Edge/Fog Computing2 references5 citations

TL;DR

This paper provides a comprehensive review of embedded development boards optimized for Edge-AI workloads, evaluating their hardware capabilities, AI inference performance, and suitability for real-time applications. It evaluates platforms like NVIDIA Jetson, Google Coral, and Raspberry Pi, identifying key trade-offs in performance, power efficiency, and cost for deploying machine learning models at the network edge.

ABSTRACT

The use of Deep Learning and Machine Learning is becoming pervasive day by day which is opening doors to new opportunities in every aspect of technology. Its application Ranges from Health-care to Self-driving Cars, Home Automation to Smart-agriculture, and Industry 4.0. Traditionally the majority of the processing for IoT applications is being done on a central cloud but that has its issues; which include latency, security, bandwidth, and privacy, etc. It is estimated that there will be around 20 Million IoT devices by 2020 which will increase problems with sending data to the cloud and doing the processing there. A new trend of processing the data on the edge of the network is emerging. The idea is to do processing as near the point of data production as possible. Doing processing on the nodes generating the data is called Edge Computing and doing processing on a layer between the cloud and the point of data production is called Fog computing. There are no standard definitions for any of these, hence they are usually used interchangeably. In this paper, we have reviewed the development boards available for running Artificial Intelligence algorithms on the Edge

Motivation & Objective

• To evaluate and compare embedded development boards suitable for deploying AI models at the edge.
• To identify hardware platforms that balance computational performance, power efficiency, and cost for real-world AI applications.
• To guide researchers and developers in selecting optimal boards for edge-based machine learning workloads.
• To analyze the feasibility of running deep learning models directly on edge devices, reducing reliance on cloud processing.
• To provide a decision framework based on technical specifications and benchmark results for edge AI deployment.

Proposed method

• Systematic review of commercially available embedded development boards with AI acceleration capabilities.
• Evaluation of key hardware components including CPU, GPU, NPU, and memory bandwidth for AI inference workloads.
• Benchmarking of model inference performance using standard AI workloads such as MobileNet, SSD, and YOLO.
• Analysis of power consumption and thermal characteristics under sustained AI processing loads.
• Comparison of software support, framework compatibility (e.g., TensorFlow Lite, PyTorch), and development ecosystem.
• Categorization of boards based on performance tiers and target application domains (e.g., robotics, smart sensors, industrial IoT).

Experimental results

Research questions

• RQ1 Which embedded development boards offer the best performance-to-power ratio for running deep learning inference at the edge?
• RQ2 How do different AI accelerators (e.g., NPU, TPU) impact inference speed and energy efficiency on edge boards?
• RQ3 What are the key hardware and software limitations affecting the deployment of real-time AI models on low-power boards?
• RQ4 How do popular edge AI platforms compare in terms of developer accessibility, framework support, and scalability?
• RQ5 What trade-offs exist between cost, performance, and energy efficiency in edge AI hardware platforms?

Key findings

• NVIDIA Jetson series offers the highest inference performance with dedicated AI accelerators, suitable for complex computer vision tasks.
• Google Coral Edge TPU boards deliver strong performance per watt, ideal for low-power, real-time edge AI applications.
• Raspberry Pi 4 with a USB-based TPU accelerator provides a cost-effective entry point but lags in raw performance and throughput.
• Power efficiency varies significantly across platforms, with TPU-optimized boards achieving up to 3–5× better performance per watt than CPU-only alternatives.
• Software stack maturity and framework compatibility (especially for TensorFlow Lite) are critical differentiators in deployment feasibility.
• No single board excels across all metrics; selection depends on application-specific needs such as latency, power budget, and cost constraints.