Skip to content

Raspberry Pi 5

The EdgeFirst Profiler runs on the Raspberry Pi 5 (aarch64). The CPU baseline uses ONNX Runtime; optional accelerators include the Hailo-8L M.2 module — the same M.2 form factor Raspberry Pi's AI Kit ships with — accessed through HailoRT.

For a guided platform tour see the Raspberry Pi Quick Start. This page covers only the profiler-specific setup.

Prerequisites

  • Raspberry Pi OS 64-bit (Bookworm or later)
  • libonnxruntime.so (sudo apt install libonnxruntime or pip install onnxruntime)
  • For the Hailo accelerator: HailoRT installed and libhailort.so on the loader path — see the Hailo install guide

Install the profiler

pip install edgefirst-profiler

curl -fsSL https://raw.githubusercontent.com/EdgeFirstAI/profiler-cli/main/install.sh | bash

Confirm:

edgefirst-profiler --version

CPU baseline (ONNX Runtime)

The default --provider cpu is the most predictable baseline. CPU inference on the RPi5 makes the capture and preprocess stages a meaningful fraction of total wall time — the Studio trace view will show whether the bottleneck is the model or the pipeline around it.

Real-time scheduling and sudo

Real-time inference scheduling (SCHED_FIFO) requires elevated privileges on the Raspberry Pi, and it gives lower and more consistent inter-inference latency. The TUI asks before validation starts whether to run under sudo or to continue at normal priority; passwordless sudo skips the password prompt. If a run fails partway on a privilege error instead — for example NPU or GPU zero-copy access being denied — the dashboard explains that elevated privileges are required and offers to re-run under sudo.

Either way, trace and prediction files are handed back to your user afterward so the EdgeFirst Studio upload still works; the explicit --output-owner <uid:gid|username> flag (or the EDGEFIRST_OUTPUT_OWNER environment variable) overrides the target owner for scripted or root workflows. Each run also records its privilege diagnostics in profiler.log — the user it ran as, whether it was elevated via sudo, and whether inference obtained real-time scheduling priority or fell back to normal — so you can confirm from the log alone that an elevated retry took effect.

Hailo-8 / 8L accelerator

When a Hailo M.2 module is present and HailoRT is installed, the profiler routes compiled .hef files through the Hailo backend automatically. The Hailo backend is described in detail on the Hailo install page, including the libhailort_profiler shim that exposes per-context timing.

On the RPi5 with a Hailo-8L, the measured auto inference depth is 4, matching the HailoRT async scheduler's driver-reported depth. The Hailo-only --batch-size flag exists but device batching is still under development: a value above 1 is accepted with a warning and the run proceeds per-frame, which is the optimized low-latency default.

Thermal considerations

The RPi5 will throttle under sustained inference if the SoC reaches ~85 °C. The profiler samples thermal-zone temperatures and reports them as system-metric counters; if the temperature climbs steadily during a long run, look at the per-iteration latency in the dashboard — a sudden bump usually corresponds to a thermal throttle event. On the Pi 5 the readout also includes the RP1 I/O controller's temperature, captured through its hwmon sensor alongside the SoC thermal zones.

A small heatsink and fan is enough to keep the SoC out of throttle territory during validation runs.

Verifying the install

edgefirst-profiler login
edgefirst-profiler              # opens TUI on F1 Help

Then run a validation session — see Validation from Studio or Validation from the Profiler.