Power trade-off between on-device AR inference and offloading

Determine, for augmented reality applications on smartphones that perform deep neural network analytics on each camera frame, whether executing inference locally on the mobile GPU or offloading inference to edge GPU servers minimizes mobile-device battery consumption, explicitly accounting for the wireless network interface energy required to upload camera frames.

Background

The paper discusses a case paper on augmented reality (AR) workloads, which can either run locally on the mobile GPU or be offloaded to edge GPU servers. Offloading requires transmitting large camera frames over Wi‑Fi or cellular networks, which incurs additional energy on the mobile device’s wireless interface.

The authors note that, from a power perspective, it is not clear a priori which approach—local processing or offloading—will consume less battery on the mobile device. They propose that accurate power models for both the mobile GPU and the wireless NIC can help developers make this decision for specific workloads and devices, and they provide example GPU power estimates for common AR tasks.