Jetson Clock Customizer
Lock or cap Jetson CPU, GPU, and EMC clock behavior before flashing a BSP image.
The skill restricts writes to the source overlay, forbids direct edits to the read-only BSP, and discloses sudo, flashing, power increase, package overwrite, and reversal paths; there is no credential handling or exfiltration behavior. Points are deducted because it changes boot-level DTB/rootfs state and flashes hardware without an explicit per-operation confirmation requirement, thermal/power safety thresholds, or a complete failure-recovery plan.
Prerequisites, SKU dispatch, SoC branching, idempotency, re-enable steps, and troubleshooting are described concretely, with a referenced canonical protocol. Static review cannot verify commands, paths, dtc output, or deployment behavior; the benchmark has only two positive tasks and is not an independently reproducible test suite, so the score remains below the static ceiling.
Triggers, supported operations, exclusions, T23x/T26x differences, and the two-layer clock ceiling are clearly stated, and the core function does not depend on overseas services. Points are deducted because the scope is limited to Jetson BSP pre-flash workflows, with little evidence for broader hardware/version boundaries, Chinese-language interaction, or finer anti-mis-trigger conditions.
The documentation is layered into purpose, prerequisites, operations, recipes, limitations, troubleshooting, and references, and includes a version, Apache-2.0 license, owner/maintenance clues, and benchmark record. Points are deducted for the missing recommended Examples section, nonconforming author format, lack of a clear changelog and maintenance/update path, and repository license metadata marked NOASSERTION.
The skill provides directly usable BPMP DTB, nvpower.sh, nvpmodel, and deployment procedures, including MAXN and multi-node GPU cases. Static review cannot confirm device outcomes; the benchmark covers only two positive cases with limited effectiveness results and depends on sibling skills and manual checks, so completeness and comparative-benefit points are deducted.
The files contain concrete commands, configuration mappings, commit conventions, expected failures, and evaluation ground truth, providing moderate auditability. Points are deducted because there is no inspectable CI/test suite, independent corroboration, or captured execution output; benchmark claims cannot replace independent static reproduction.
- Execution changes boot-level BPMP DTB/rootfs state and may flash the device; obtain explicit user confirmation before writes or flashing and preserve a recoverable pristine image and rollback path.
- MAXN and disabling EMC DVFS can materially increase power, temperature, and hardware stress; the skill provides no safety thresholds or monitoring procedure.
- Resolve the DTB using the active profile, SKU/FAB dispatch, and SoC detection; a static BPFDTB_FILE lookup or omitted T23x GPU partitions can make the configuration ineffective or undefined.
- nvidia-l4t-init upgrades can overwrite nvpower.sh; long-lived deployments should validate the proposed systemd drop-in or helper-script alternative.
What does this skill do, and when should you use it?
This skill is for developers customizing an NVIDIA Jetson BSP image. It edits the BPMP DTB and boot-time nvpower.sh to configure clock ceilings, EMC DVFS, and cpufreq/devfreq governors. Changes are staged in a source overlay tracker and deployed through the promote-and-flash workflow. Live runtime tuning and nvpmodel power-mode edits are out of scope.
Checks the active target profile, extracted BSP image, and initialized source overlay tracker; resolves the SKU-correct BPFDTB_FILE from the active flash configuration; decompiles and recompiles the BPMP DTB with dtc; edits max-rate-custom, bwmgr, cactmon, and the SoC-specific osp-controller path; edits Linux_for_Tegra/rootfs/etc/systemd/nvpower.sh for governors or device rate bounds; creates git commits and directs deployment through /jetson-promote-image and /jetson-flash-image.
- A Jetson BSP developer needs to lock a CPU or GPU to a specified clock rate before flashing.
- A performance tester needs to disable EMC DVFS so EMC remains at its initialization rate.
- A systems engineer needs to change CPU/GPU governors or apply device minimum and maximum rate bounds.
- A stress-testing workflow needs BPMP DTB, nvpower.sh, and MAXN nvpmodel changes combined.
What are this skill's strengths and limitations?
- Covers both BPMP DTB and nvpower.sh clock-control layers.
- Documents T23x/T26x EMC DVFS differences, SKU-dispatched configurations, and troubleshooting cases.
- Preserves changes through the overlay tracker, commit conventions, and promote/flash workflow.
- The repository README states that published skills carry NVIDIA signatures, evaluations, and security scanning.
- Supports image-build and flashing workflows, not live-target clock tuning.
- max-rate-custom can only be set below max-rate-maxn; it cannot raise the silicon ceiling.
- nvpower.sh is managed by the nvidia-l4t-init package, so upgrades may overwrite in-place edits.
- The supplied source does not provide independent test results or a verified-platform list for this skill.
How do you install this skill?
Run: npx skills add nvidia/skills --skill jetson-customize-clocks --yes. Reload skills in an Agent Skills-compatible client before use.
How do you use this skill?
Give the agent a concrete request such as “lock the Jetson GPU frequency” or “disable EMC DVFS.” Before editing, provide an active target profile, an extracted BSP image, and an initialized source overlay tracker. Deploy completed changes with /jetson-promote-image → /jetson-flash-image; route live tuning and nvpmodel edits to the relevant skills.
How does this skill compare with similar options?
The source explicitly assigns nvpmodel power-mode edits to the sibling /jetson-customize-nvpmodel skill and deployment to /jetson-promote-image and /jetson-flash-image. No other alternative tool is specified.