New power model for federated learning on mobile devices slashes errors from 959% to under 10%
Federated learning on phones just got 1.4x more energy efficient with a new CPU power estimation method.
Get AI news that actually matters
One email a day. Zero fluff. Join 10,000+ professionals.
Estimating CPU power on heterogeneous ARM-based mobile devices is notoriously difficult because commodity Android phones hide voltage domain details. Most energy-aware federated learning (FL) frameworks resort to simplified approximate power models rather than the more accurate analytical CMOS-based model. This approximation can lead to wildly inaccurate energy estimates, causing suboptimal training decisions and wasted battery life.
To address this, the team introduces a rail-to-cluster mapping technique that retrieves cluster-level supply voltage from two commodity Android devices. Their analytical model predicts CPU power with errors below 10%, while the approximate model can incur errors up to 959%. When tested with the AnycostFL framework, the analytical model achieves the same 80% model accuracy while using 1.4x less energy. This work makes analytical power models practical on heterogeneous multi-cluster ARM-based mobile SoCs without any additional hardware or external power measurement tools, opening the door to far more efficient on-device AI training.
- Proposed analytical power model achieves <10% error vs. approximate models that can reach 959% error on ARM mobile devices.
- Rail-to-cluster mapping technique extracts voltage data without extra hardware, enabling reproducible power estimation.
- In AnycostFL, the model delivers same 80% accuracy while consuming 1.4x less energy, improving battery efficiency in federated learning.
Why It Matters
Accurate power modeling unlocks more energy-efficient federated learning on everyday phones, reducing battery drain without sacrificing AI performance.