Build an auditable inference path for ArmSoM RK3576 and RK3588 devices while separating vendor claims from measured results.
What this article helps you decide
A deployable model is more than an RKNN file: preprocessing, quantization, tensor layout, runtime version, and validation samples must be versioned together. This guide is an engineering workflow, not a substitute for the schematic, release notes, or safety instructions for the exact ArmSoM product and revision.
Platform context
Both product families advertise NPU capability; actual latency and accuracy require a declared model, input shape, runtime, clocks, thermal state, and measurement method.
The practical rule is to keep the board, BSP/image version, peripheral revision, power arrangement, and reproducible commands together in one project record. That makes a field issue diagnosable and prevents a working lab configuration from becoming an undocumented dependency.
Reproducible starting procedure
Use the matching RKNN runtime release and compare outputs against a fixed reference set.
python3 - <<'PY'
# Pseudocode: record metadata next to each model artifact
print({"model":"model.rknn", "input":"1x3x224x224", "runtime":"verify-on-target"})
PY
Record the command output, kernel release, image release, board revision, and any attached peripheral identifiers. When a command writes media, changes boot configuration, or controls power, stop and consult the board-specific manual before proceeding.
Validation checklist
- Confirm the exact ArmSoM board, revision, memory/storage SKU, and power method.
- Capture
uname -a, image/build version, and the full serial boot log. - Verify the expected hardware node or connector before tuning performance.
- Change one variable at a time and keep the before/after evidence.
- State workload, environment, software versions, and method with any result.
Troubleshooting
Symptom: the expected device or behavior is absent. First compare the running image and device tree to the product documentation, then check power, cable direction, connector seating, and boot logs.
Symptom: the command exists but gives an unexpected result. Do not publish throughput from a host conversion run. Measure on the target with warm-up, repetitions, percentile reporting, and stated thermal/power conditions.
Symptom: a configuration appears to work once but not after reboot. Check whether the change belongs in the image, boot configuration, device tree, service definition, or a transient runtime command. Preserve the serial log and a minimal reproduction.
Publishing notes
This article intentionally contains no fabricated benchmark, compatibility claim, or laboratory measurement. Add measured results only with the test unit, software revision, workload, ambient conditions, repetitions, and raw evidence.
References
- ArmSoM Open Source Hardware Developer Center
- ArmSoM Sige5 RK3576 documentation
- ArmSoM Sige7 RK3588 documentation
- ArmSoM AIM5 RK3576 documentation
- ArmSoM AIM7 RK3588 documentation
- Linux kernel DeviceTree documentation
- Rockchip RK3588 brief datasheet
Pre-publication check
- Verify each product capability and pin/interface statement against the current board documentation.
- Replace generic commands with the tested image/release-specific command where needed.
- Confirm screenshots, diagrams, and alt text match the final board and workflow.
- Add a canonical URL and author/editor attribution in front matter.
- Have a technical owner review safety, flashing, electrical, and availability statements.