Speaker

A Systematic Approach for Performance Evaluation of Edge-AI Hardware and Software Applications

Edge-AI has emerged as a promising paradigm for deploying artificial intelligence algorithms directly on edge devices,
offering numerous advantages including reduced latency, enhanced privacy, and bandwidth efficiency.
However, its adoption is accompanied by a set of challenges, particularly concerning the performance limitations of
small hardware devices such as the MAX78000FTHR, Raspberry Pi, Jetson Nano, Orin, or Xavier. Our analysis focuses on
the performance of Edge-AI on these platforms, comparing a diverse set of model architectures including Fully Connected,
LSTM, Conv1D, and Conv2D networks.
We explore the implications of utilizing model languages such as PyTorch, Keras, TensorFlow Lite, and ONNX, and analyze
their suitability for deployment in resource-constrained environments. Through comprehensive benchmarking and
experimentation, we aim to provide insights into the suitability of these model architectures for deployment on
resource-constrained edge devices. By examining factors such as inference speed, memory usage, and power efficiency,
we seek to identify the most effective model configurations for Edge-AI applications in real-world scenarios. This enables
us to assist customers in informed decisions-making and finding suitable solutions in this rapidly expanding field.

Our work aims to develop a framework that simplifies the analysis of hardware and software properties,
relevant to Edge-AI environments. It helps to examine various hardware and software configurations to
effectively navigate the complexities of deploying AI at the edge.