Assessment of Embedded AI Solutions with the Green Software Measurement Model
Year: 2025
Type: conference-paper
Abstract: When developing embedded systems running AI-based software, different infrastructures with different trade-offs are possible. To support the decision for a specific infrastructure, we use the Green Software Measurement Model and electric power measurements to determine the energy efficiency of two neural networks across three target platforms. The dedicated low-power ML accelerators show a much higher energy efficiency than the high-performance GPU. But with their limited memory, they may be unable to run larger neural networks. Besides, the ultra-low-power Coral accelerator falls behind when using non-optimized models. We make the full data and data evaluation publicly available.
Authors: Christoph Bockisch, Hartmut Weber, Dennis M. Pöpperl, Severin Stahl
Language: English
Topic: Green IT and Sustainability
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4411447354