Predicting Delay, Power, and Energy of Edge AI Inference with Ridge Regression
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
As AI becomes central to consumer devices, edge inference must meet strict latency and energy constraints under limited compute and thermal budgets. We present a regression-based framework for predicting four inference metrics—Delay, Power, Energy, and Energy–Delay Product (EDP)—on the NVIDIA Jetson Orin Nano. The predictor leverages structural features (FLOPs, parameters, depth, Conv/MLP ratio) and run-time settings (GPU frequency, batch size). Delay and Energy are modeled with log-linear Ridge regression, while Power is refined from a physics-based approximation (P ≈ E/T ) with residual correction. Evaluation under random, architecture, and runtime holdouts demonstrates high accuracy for Delay and Energy, robust Power prediction across architectures, and moderate sensitivity to runtime shifts. The framework provides interpretable formulas that enable rapid design-space exploration for consumer electronics such as smart cameras, wearables, XR devices, and healthcare systems.
Fields of Study & Subjects
Advanced Neural Network Applications
IoT and Edge/Fog Computing
Parallel Computing and Optimization Techniques
Publication details
- Journal / source
- International Conference on Consumer Electronics
- Pages
- 1–6
- Publication date
- 3 February 2026
- Language
- EN