Project Overview

This project evaluates the performance and efficiency of artificial intelligence inference across heterogeneous computing platforms, with particular focus on Edge AI deployment. Edge platforms differ significantly in processing capability, memory, supported inference runtimes, and resource constraints, making it difficult to determine their suitability for AI workloads from hardware specifications alone.

The study evaluates two edge platforms, the Raspberry Pi 5 and Samsung Galaxy S26 Ultra, using an AMD Ryzen 7 6800H CPU and NVIDIA GeForce RTX 4090 GPU as higher-performance reference platforms. The evaluation combines MLPerf Tiny Image Classification, Keyword Spotting, and Visual Wake Words workloads with local Gemma 3 1B inference across five language-based workload categories.

Performance is assessed using measurements including inference time, throughput, classification accuracy, F1-score, ROC-AUC, token-processing rates, and estimated carbon emissions. The project also examines the effect of INT8 model quantisation on the Raspberry Pi 5. The results show that platform suitability depends not only on hardware capability, but also on the workload, model representation, inference runtime, and execution environment. The project therefore provides a comparative basis for understanding the trade-offs involved when selecting platforms for Edge AI deployment.