Is Hardware Offload Enough? Energy, Timing, and Scheduling of a Video Analytics Pipeline on the Raspberry Pi 4

Single-board computers ship with hardware video codecs, and the usual advice for a video pipeline with deadlines is to use them. We test that advice on a Raspberry Pi 4 running a 720p decode-detect-overlay-encode pipeline with a person detector on every tenth frame. Moving decoding and encoding to the hardware codec raises sustained throughput from 18 to 30 frames/s. At equal work (15 frames/s) it cuts CPU time per frame by 52%, but outlet energy per frame, measured in one session, by only about 16%. Lower CPU time did not mean lower latency: in the way this application collects encoder output, each hardware-encoded packet was retrieved one or two frames after submission, so at 15 frames/s the median time from a frame's release to its encoded packet was 148 ms against 72 ms in software. With detection moved to a separate worker on one core at 15 frames/s, every hardware-mode frame-task and detection job met its CPU deadline under rate-monotonic priorities, with and without competing load, below an idealized response-time estimate; under the default Linux policy, competing load made every frame-task job late. Smaller findings concern slow CPU access to decoder buffers and the frequency governor.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-12
DOI
https://doi.org/10.5281/zenodo.24166002
Primary Topic
Real-Time Systems Scheduling
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Is Hardware Offload Enough? Energy, Timing, and Scheduling of a Video Analytics Pipeline on the Raspberry Pi 4

Yazeed Alzughaibi
Zenodo (CERN European Organization for Nuclear Research)
Real-Time Systems Scheduling
preprint

Is Hardware Offload Enough? Energy, Timing, and Scheduling of a Video Analytics Pipeline on the Raspberry Pi 4

Yazeed Alzughaibi
preprint en

Abstract

Single-board computers ship with hardware video codecs, and the usual advice for a video pipeline with deadlines is to use them. We test that advice on a Raspberry Pi 4 running a 720p decode-detect-overlay-encode pipeline with a person detector on every tenth frame. Moving decoding and encoding to the hardware codec raises sustained throughput from 18 to 30 frames/s. At equal work (15 frames/s) it cuts CPU time per frame by 52%, but outlet energy per frame, measured in one session, by only about 16%. Lower CPU time did not mean lower latency: in the way this application collects encoder output, each hardware-encoded packet was retrieved one or two frames after submission, so at 15 frames/s the median time from a frame's release to its encoded packet was 148 ms against 72 ms in software. With detection moved to a separate worker on one core at 15 frames/s, every hardware-mode frame-task and detection job met its CPU deadline under rate-monotonic priorities, with and without competing load, below an idealized response-time estimate; under the default Linux policy, competing load made every frame-task job late. Smaller findings concern slow CPU access to decoder buffers and the frequency governor.

Zenodo (CERN European Organization for Nuclear Research)
King Abdulaziz University (SA)
Real-Time Systems Scheduling
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon