jetson-customize-clocks

द्वारा nvidia

Use to lock/cap Jetson CPU/GPU/EMC clocks, toggle EMC/CPU DVFS, or change cpufreq governors by editing BPMP DTB and nvpower.sh pre-flash. Do NOT use for live…

npx skills add https://github.com/nvidia/skills --skill jetson-customize-clocks

Customize Clocks

Purpose

Customize CPU, GPU, and EMC clock behavior on a Jetson target by editing files under Linux_for_Tegra/ before flashing the image. Two layers are in scope:

  • The BPMP DTB at Linux_for_Tegra/bootloader/<BPFDTB_FILE> — per-clock max-rate-custom ceilings, plus the EMC DVFS gate (bwmgr + cactmon on all SoCs; osp-controller on T26x only).
  • nvpower.sh at Linux_for_Tegra/rootfs/etc/systemd/nvpower.sh — cpufreq / devfreq governors and (optionally) per-device min / max / static rates written to sysfs at boot.

Common triggers: "lock CPU/GPU/EMC frequency", "pin GPU to Fmax", "pin EMC to MAXN", "disable/enable EMC DVFS", "disable/enable CPU DVFS", "set CPU/GPU max rate", "change cpufreq governor".

Out of scope: runtime clock tuning on a live target (no flash step), nvpmodel power-mode edits (use the sibling skill /jetson-customize-nvpmodel), and silicon-ceiling overrides (max-rate-maxn is read-only).

Prerequisites

Resolve the active profile per ../../context/target-platform-contract.md. Refuse and route in these cases:

ConditionRefuse with
No active profile, or active: NARoute to /jetson-set-target or /jetson-init-target.
Profile lacks bsp_image: blockRoute to /jetson-init-image.
<bsp_image.root_path>/Linux_for_Tegra/ missingRoute to /jetson-init-image.
<source.root_path>/Linux_for_Tegra/ missing or not a git repoRoute to /jetson-init-source.

Resolve paths:

  • <bsp_image.root_path> from bsp_image.root_path: if present, else <workspace>/Image.
  • <source.root_path> from source.root_path: if present, else <workspace>/Source.

<bsp_image.root_path> is read-only for this skill; every write (Operation 1's BPMP DTB and Operation 2's nvpower.sh) lands under <source.root_path> (the overlay tracker). This is the workflow invariant in ../../context/bsp-customization-workflow.md#workflow-invariants — hand-editing upstream silently destroys the diff trail and makes /jetson-promote-image a noop.

Instructions

  1. Resolve the prerequisites above (active profile, BSP image extracted, source overlay tracker initialized).
  2. Pick the operation from the table below.
  3. Follow the linked procedure section — Operation 1 (BPMP DTB), Operation 2 (nvpower.sh), or the MAXN recipe for both.
  4. Commit the edit inside the overlay tracker per each Operation's commit convention.
  5. Deploy with /jetson-promote-image/jetson-flash-image. The new BPMP DTB and nvpower.sh take effect on the next boot.

Supported operations

OperationWhere the edit livesProcedure section
Lock a CPU / GPU clock to a specific rateBPMP DTB max-rate-custom on the clock node + nvpower.sh governor performance"Content edit: max-rate-custom" + "Pick the edit"
Lock EMC at its init rate (disable EMC DVFS)BPMP DTB: bwmgr.enabled = 0, cactmon.enabled = 0, plus /delete-node/ osp-controller on T26x only"Content edit: EMC DVFS disable / enable"
Re-enable EMC DVFSBPMP DTB: bwmgr.enabled = 1, cactmon.enabled = 1, restore osp-controller on T26x"Content edit: EMC DVFS disable / enable"
Pin everything to MAXN for stress runsCombine the above + nvpmodel MAXN as boot defaultsee Recipe
Lower a clock's hard ceiling without lockingBPMP DTB max-rate-custom only"Content edit: max-rate-custom"
Bound a device's rate without pinningnvpower.sh min/max via sysfs"Pick the edit"

Operation 1 — BPMP DTB edits

Follow the BPMP-DTB customization protocol in ../../references/bsp-customization-bpmp-dtb.md. The protocol owns the mechanics — pristine import on first touch, dtc decompile, recompile, sanity-check, commit. This skill supplies only the clock-specific content (which nodes and properties to edit during the protocol's "Edit the DTS" step).

The edited .dtb lands in the <source.root_path>/Linux_for_Tegra/ overlay tracker. /jetson-promote-image's channel A walks the tracker and copies the file into bsp_image. Do not edit <bsp_image.root_path>/Linux_for_Tegra/bootloader/<bpmp-dtb> directly — that's the promote output, not an input.

Resolve the SKU-correct BPMP DTB

Per the protocol's "Resolving the active BPMP DTB" section, read BPFDTB_FILE from the active flash conf. For the common Thor / single-SKU conf shapes this is the static BPFDTB_FILE=... line in the per-board .conf and the value is authoritative as-is.

For SKU-multiplexed conf shapes (Orin AGX devkit conf chain that selects a different BPMP DTB per board_sku/board_FAB via update_flash_args_common — see ../../context/bsp-customization-software-layers.md#per-board-conf-dispatch--update_flash_args_common), walk the dispatch chain with board_sku=<module.sku> and board_FAB=<module.revision or empty> from the active profile, and read BPFDTB_FILE from the dispatch output — not from the static line of the per-board .conf. Static and dispatched values match for non-multiplexed confs; the dispatch is mandatory only when the conf chain conditionally overrides BPFDTB_FILE.

List effective max rates (inspection)

Inspect both layers of the runtime ceiling — see references/clock-control-model.md#effective-runtime-ceiling — before deciding on a max-rate-custom value.

Inspection cookbook (BPMP-side decompile + grep; nvpmodel-side awk over the boot default mode) is in references/bpmp-dtb-clock-edits.md#inspection-cookbook.

For the nvpmodel layer see /jetson-customize-nvpmodel.

This step does not mutate state — it's a precondition for sizing the edit in the "Content edit: max-rate-custom on a named clock node" step.

Content edit: max-rate-custom on a named clock node

During the "Edit the DTS" step of the protocol, modify the property inside the named clock node — never lateinit. max-rate-custom must be strictly below the clock's hard cap (max-rate-maxn if defined, otherwise the live max_rate from a running target of the same chip / SKU).

DTS edit form, semantics, and the nvpmodel ↔ BPMP clock-node mapping live in references/bpmp-dtb-clock-edits.md.

Then hand control back to the protocol — its "Recompile", "Sanity-check the recompiled blob", "Stage in the overlay tracker", and "Cleanup" steps cover the rest. Commit-message convention per the protocol: <BPMP_BASENAME>: jetson-customize-clocks — <clock-node> max-rate-custom = <value>.

Content edit: EMC DVFS disable / enable

Default behavior (EMC DVFS on) requires no edit. Disabling EMC DVFS is a multi-node edit applied inside the same "Edit the DTS" step of the protocol, not a bwmgr toggle:

#EditScope
1bwmgr.enabled = <0x00>All SoCs, mandatory
2cactmon.enabled = <0x00>All SoCs, mandatory
3/delete-node/ osp-controllerT26x (Thor) mandatory — T23x (Orin) has no such node, skip

Detection: dtc -I dtb -O dts <bpmp-dtb> | grep -c osp-controller — zero hits ⇒ T23x path. Full DTS snippets, the surviving-paths failure modes, and the re-enable procedure are in references/emc-dvfs-disable.md.

Apply the protocol's "Recompile" through "Cleanup" steps once the multi-node edit is in place. Commit-message convention: <BPMP_BASENAME>: jetson-customize-clocks — EMC DVFS disable (bwmgr + cactmon[+ osp-controller]).

Disabling raises idle power; intended for stress / performance tests, not production rootfs.

Re-run + idempotency

Per the protocol's "Re-runnability" section, re-running this skill with the same target value produces a no-op commit. Re- running with a different value rewrites the same property — git log -- $BPMP_REL shows the per-run history. To return a clock to its max-rate-maxn ceiling, edit the DTS to remove the max-rate-custom line and recompile.

Operation 2 — nvpower.sh edits

Edits nvpower.sh, which runs at boot via nvpower.service to set cpufreq / devfreq governors and rates.

The per-script file

The script this Operation edits has the relative path:

Linux_for_Tegra/rootfs/etc/systemd/nvpower.sh

It lives in two roots; the Operation walks both:

RoleLocationSkill writes?
Detection + pristine source<bsp_image.root_path>/Linux_for_Tegra/rootfs/etc/systemd/no — read-only
Overlay edit target + git commit<source.root_path>/Linux_for_Tegra/rootfs/etc/systemd/yes

Subsequent sub-steps refer to the per-script file to mean the overlay copy under <source.root_path>. The <bsp_image.root_path> copy is read once during the pristine-import step below, then never touched again.

Overlay edit recipe (apply before editing nvpower.sh)

Follow the canonical Off-skill edits recipe in the workflow doc — pristine import + customization commit pair, both gated by the preview gate. nvpower.sh is a single file with no propagation set; one pristine commit + one customization commit covers the entire change.

Concrete substitutions for this skill:

  • <rel>/<file> is rootfs/etc/systemd/nvpower.sh.
  • Suggested pristine-import message: import pristine: rootfs/etc/systemd/nvpower.sh, body Source: <bsp_image.root_path>/Linux_for_Tegra/ (BSP <bsp_image.version>).
  • Suggested customization-commit header: jetson-customize-clocks: nvpower.sh <summary>, body lines like set_cpufreq_governor: desired_cpufreq_gov "schedutil" -> "performance".

Pick the edit

Function locations (set_cpufreq_governor, set_devfreq_governor), common-edit recipes (pin to Fmax, static rate, min/max bounds), and the nvidia-l4t-init package-upgrade caveat live in references/nvpower-sh-edits.md.

Deploy

The customization commit in the overlay tracker does not reach the device on its own. The Deploy chain:

  1. /jetson-promote-image — copies every tracked file in the overlay into <bsp_image.root_path>/Linux_for_Tegra/. Diff-aware (skip byte-identical); uses sudo cp -p for rootfs/* destinations.
  2. /jetson-flash-image — flashes the updated bsp_image to the device. nvpower.service runs the new script on the next boot.
  3. (Alternate, no flash) Copy <source.root_path>/Linux_for_Tegra/rootfs/etc/systemd/nvpower.sh directly to the running target's /etc/systemd/nvpower.sh, then sudo systemctl restart nvpower.service (or reboot).

Editing <source.root_path>/... without committing — or editing <bsp_image.root_path>/... directly — does nothing for /jetson-promote-image and is silently lost on the next /jetson-init-image re-extract.

Recipe — pin everything to MAXN for stress / performance runs

Combines Operations 1 + 2. Operation 1's BPMP edits all flow through one round of the protocol (a single decompile / multi-node edit / recompile / commit cycle — don't round-trip the protocol twice for the same .dtb):

  1. BPMP DTB (the "Content edit: max-rate-custom on a named clock node" step content): leave max-rate-custom unset on every CPU / GPU / EMC clock; remove existing max-rate-custom lines that lower the ceiling.
  2. BPMP DTB (the "Content edit: EMC DVFS disable / enable" step content): pin EMC at its init rate — bwmgr.enabled = 0, cactmon.enabled = 0, plus /delete-node/ osp-controller on T26x (skip on T23x).
  3. Apply both content edits inside one protocol "Edit the DTS" invocation, then run the remaining protocol steps (recompile, sanity-check, single customization commit covering both content edits).
  4. nvpower.sh (Operation 2): set desired_cpufreq_gov="performance" and desired_devfreq_gov="performance" unconditionally; remove the GPU/nvjpg skip in set_devfreq_governor. Applies via Operation 2's overlay edit recipe (the "Overlay edit recipe (apply before editing nvpower.sh)" step) — a separate overlay-tracker pristine + customization commit pair on the rootfs script, distinct from the BPMP-DTB protocol's commit.
  5. Set the boot-default nvpmodel mode to MAXN via /jetson-customize-nvpmodel — the per-clock nvpmodel cap clamps below max-rate-maxn regardless of BPMP DTB content.

Deploy /jetson-promote-image/jetson-flash-image picks up the new BPMP DTB (via the overlay tracker) and the edited nvpower.sh (via the same overlay tracker) on the next flash.

Limitations

  • Image-build-time only. All edits land under <source.root_path>/Linux_for_Tegra/ and reach the device only via /jetson-promote-image/jetson-flash-image. Live-target tuning is out of scope.
  • max-rate-custom only lowers the ceiling. It must be strictly below max-rate-maxn; raising the silicon cap is not supported.
  • Effective ceiling is two-layer. The runtime ceiling is min(BPMP cap, active-nvpmodel-mode cap). The nvpmodel cap is owned by /jetson-customize-nvpmodel; this skill does not edit it.
  • SoC-conditional EMC DVFS gate. Disabling EMC DVFS requires editing different node sets on T23x (bwmgr + cactmon) vs T26x (bwmgr + cactmon + delete osp-controller). Mis-detection produces undefined behavior.
  • T23x GPU cap is multi-node. The GPU clock is split across nafll_gpusys and every nafll_gpcX; the cap binds only when applied to all of them.
  • nvpower.sh is package-managed. It ships in nvidia-l4t-init; package upgrades clobber in-place edits. Long-lived setups should prefer a systemd drop-in or sibling helper.
  • ODMDATA wins. When an ODMDATA token covers a property, the token overrides direct BPMP DTS edits at flash time. Direct BPMP DTS edits are the fallback for properties no NVIDIA token reaches.
  • max-rate-maxn and lateinit are off-limits. max-rate-maxn is the silicon ceiling (read-only). lateinit is for boot-time clock init, not ceiling overrides — never touch either.
  • BPMP DTB may be SKU-multiplexed. On compound / dispatched flash confs (Orin AGX devkit chain), BPFDTB_FILE is selected by board_sku / board_FAB via update_flash_args_common. Reading the static BPFDTB_FILE= line is wrong when the chain conditionally overrides it; resolve via the dispatch instead.

Troubleshooting

ErrorCauseSolution
max-rate-custom set but clock still ramps to max-rate-maxn on T23x GPUOnly nafll_gpusys was capped; the nafll_gpcX partitions still run at max-rate-maxn and dominate the effective ceiling.Apply the same max-rate-custom to nafll_gpusys and every nafll_gpcX node enumerated by grep -nE '^\s*nafll_gpc[0-9]+\s*:' <decompiled.dts>.
EMC DVFS disable appears to apply but EMC still scales on T26xOnly bwmgr.enabled = <0x00> was set; osp-controller survives and re-issues frequency changes via the QoS path.Add edits #2 (cactmon.enabled = <0x00>) and #3 (/delete-node/ osp-controller) inside the same "Edit the DTS" step. Verify osp-controller via dtc -I dtb -O dts <bpmp-dtb> | grep -c osp-controller → expect 0.
EMC DVFS disable rejected on T23x with "node not found" for osp-controllerT23x (Orin) BPMP DTBs do not contain osp-controller; edit #3 must be skipped on T23x.Detect SoC family with the grep -c osp-controller step; only apply #3 when the count is ≥1.
BPMP refuses to load DTB after edit: max-rate-custom >= max-rate-maxnmax-rate-custom was set to or above the silicon ceiling.Lower max-rate-custom strictly below max-rate-maxn. If max-rate-maxn is absent from the node, query the live cap on a running target: cat /sys/kernel/debug/bpmp/debug/clk/<clock>/max_rate.
osp-controller re-appears after status = "disabled"status = "disabled" does not remove the node from the device tree; BPMP still walks it.Replace with /delete-node/ osp-controller; — the node must not exist for BPMP to skip the path.
Edits to nvpower.sh lost after apt upgradenvpower.sh is owned by the nvidia-l4t-init deb and gets overwritten on upgrade.For long-lived test setups, package edits into a systemd drop-in or a sibling helper file referenced by nvpower.sh, rather than editing nvpower.sh in place.
/jetson-promote-image is a no-op after editing the BPMP DTBThe edit was applied to <bsp_image.root_path>/Linux_for_Tegra/, which is /jetson-promote-image's output — not its input.Move the edit to <source.root_path>/Linux_for_Tegra/bootloader/<BPFDTB_FILE> (the overlay tracker) and commit through the BPMP-DTB protocol.
Cap appears to apply on first boot then resets after a power-mode changeThe active nvpmodel mode's per-clock cap clamps below max-rate-custom.Inspect both layers; if nvpmodel is binding, raise (or remove) the nvpmodel cap via /jetson-customize-nvpmodel. The BPMP cap alone is not the runtime ceiling.

References

nvidia की और Skills

compileiq-debug
nvidia
Use when something is wrong: Search() hangs, all evaluations return INVALID_SCORE, scores aren't improving, every config returns the same number, ptxas errors…
official
create-github-pr
nvidia
Create GitHub pull requests using the gh CLI. Use when the user wants to create a new PR, submit code for review, or open a pull request. Trigger keywords -…
official
diagnose-perf
nvidia
First-responder performance triage for Isaac Sim and Isaac Lab. Identifies bottleneck category (GPU-bound, CPU-bound, VRAM, loading) using nvidia-smi and…
official
eagle3-review-logs
nvidia
Review EAGLE3 pipeline experiment logs from the launcher's experiments/ directory. Summarizes pass/fail status for all 4 tasks, diagnoses failures with root…
official
nemoclaw-maintainer-cross-issue-sweep
nvidia
Scans other open issues to find ones a given PR may also fix or accidentally break. Outputs adjacent-fix opportunities and contradiction risks with file:line…
official
karpathy-guidelines
nvidia
सामान्य LLM कोडिंग गलतियों को कम करने के लिए व्यवहार संबंधी दिशानिर्देश। कोड लिखते, समीक्षा करते या रिफैक्टर करते समय उपयोग करें ताकि अत्यधिक जटिलता से बचा जा सके, सर्जिकल बदलाव किए जा सकें,…
official
fhir-basics
nvidia
Teaches agents how FHIR R4 APIs work, what resources are available, how to query them with search parameters, and how to correctly parse all response formats…
official
underdeclared-agent
nvidia
A helpful assistant agent
official