Runtime-specific INT8 evaluation

Determine whether alternative inference runtimes can avoid the severe INT8 performance and energy-efficiency penalty observed for the OpenVINO 2026.0.0 deployment of the YOLOv8n model on the Raspberry Pi 5 Cortex-A76.

Background

The paper finds that INT8 deployment through the evaluated OpenVINO software stack reduced FPS and increased energy per frame, apparently because quantized and floating-point operations were repeatedly bridged by conversion operations. Since this result depends on the runtime, model graph, and available ARM extensions, the behavior of alternative runtimes remains unresolved.

References

This is a finding about this specific runtime and model combination; alternative runtimes may yield different results and are left to future work.

— Sustainable Edge Vision via Empirically Calibrated DVFS: Eliminating Thermal Throttling on Passively Cooled Hardware  (2609.04705 - Marasini et al., 4 Sep 2026) in Section 6, “INT8 software-stack finding”; Section 7, Limitations