Panel Power and Hardware

Use Power and hardware (/rust/power-hardware) to review power profiles, live CPU/GPU readings, heat and benchmark evidence before deciding between lower sustained watts and higher performance. Power options (/rust/power-hardware/options), GPU telemetry and caps (/rust/power-hardware/gpus) and GPU tuning and benchmarks (/rust/power-hardware/tuning) replace the three classic tabs with native forms. The earlier GPU & Power page remains a separate diagnostic view during the transition. The System hardware page (/rust/system/hardware) contains read-only inventory.

The Power and hardware page displays Current evidence (CPU/temperature sample, GPU limit and device status, power profiles and drift, tuning and completed runs), Guided task forms for administrators, Controls, effects and registered routes, and More hardware controls. Sensor values come from recorded samples: missing values are not zero and detected devices are not automatically supported for execution. A cap is a sustained power limit, not protection against a momentary current spike.

Native pages and safe workflow

The main page links to three native pages; use its guided tasks when changing host performance or verifying a cap with a before-and-after benchmark. Administrators see server-rendered forms that post to the same privilege-owning services as the former classic tabs. Viewers see evidence, explanations and navigation, but not the privileged forms. Every form label has a help marker. A successful form reply includes the full owner response rather than guessing that hardware was updated.

1. Open Power options. Expand Power state and equipment, Budget and UPS evidence, Boot and stability evidence, and Native watchdog state to inspect live device support. For manual edits, type a new value next to a current setting and press Apply selected live values; leave all unchanged rows blank. Use Save selected values to create a custom profile from selected entries. Capture all live settings saves a named snapshot. Each profile exposes Apply now, Use at boot, and, for custom profiles, Delete custom profile. A boot profile can be cleared with Stop applying profile at boot. Reapply only drifted settings affects drifted values and Save automatic reapply policy changes drift behavior. High-impact boot persistence has its own force checkbox.
2. In Power options, set the PSU, UPS and platform estimates under Save power and UPS budget. If configured, enter peer/UPS endpoint entries. Use Annotate boot to distinguish manually interrupted boots from suspected crashes. Request watchdog policy lets an administrator request arm or disarm; success must be confirmed by the hardware owner. Recheck power owner only refreshes readiness.
3. For machine restart or poweroff, choose Review restart or Review shutdown. The preview names running work and returns a short-lived, signed review. No restart or shutdown occurs at this stage. Inspect it, type the exact issued phrase, acknowledge interruption, tick confirmation and select Confirm reviewed host action. If the review expired or the action changed, request a fresh review. Do not confuse this with an inference-instance restart.
4. Open GPU telemetry and caps. Expand GPU inventory, GPU runtime, memory, process residency and thermals and Device-bound power limits and saved policy. A supported writable device shows Set GPU cap now with its device-specific watt range. Forget saved cap only removes automatic reapplication intent without restoring the live cap. Use Restore exact previous cap on the Overview guide when actual hardware restoration is required.
5. Open GPU tuning and benchmarks. Expand the GPU settings and benchmark evidence before editing clocks, memory, voltage or fan. Enter selected values and use Apply selected tuning values; use Save selected GPU settings as profile, Capture current GPU profile, or Restore recorded boot defaults for profile operations. For a trial run use Apply and benchmark with rollback, which asks for a running instance and consistent benchmark parameters. Profiles expose Apply tuning profile, Use GPU profile at boot, Apply profile and test with rollback, Delete GPU profile and Stop using GPU profile at boot. Changing clocks or fans can increase heat, power and reset risk.
6. For a baseline, use Start GPU benchmark with an actual running instance, prompt tokens, generated tokens, repetitions, mode and label. Check the observed result before comparing. Stop current GPU benchmark and Delete finished benchmark record affect the active run and saved record respectively. Live benchmark status updates without a full page reload while the page is visible; a failed poll says unavailable rather than reporting false success.
7. Back up GPU firmware (read only) and Inspect local ROM without flashing do not flash the card. A saved firmware image can be retrieved from Download ... (read-only). A ROM check may advise against flashing; do not treat inspection as authorization to install it.

All the owner routes shown on the native pages remain subject to authentication, authorization and their own device-specific checks. Any failed readback or unrecognized device must be investigated before the next action. No GPU or UPS sensor data is synthesized for display.

Guided task: cap a GPU and measure the speed cost

1. In GPU telemetry and caps, identify a supported writable GPU, its exact PCI address and reported min/max watts. Record live cap, saved cap and In effect. A missing bound readback blocks a verified cap change.
2. On a running, idle GPU instance, open GPU tuning and benchmarks and start a quick benchmark with Start GPU benchmark. Keep the instance, prompt depth, predicted tokens and repetitions fixed. Wait for a finished run; note Decode t/s, Mean W and Peak junction. Do not infer a baseline from a still-running job.
3. Stop the inference instance on Instances (/rust/instances). GPU cap writes refuse active inference. Under Guided task: cap a GPU and measure the speed cost, enter PCI address and Whole watts within the card's range, tick the confirmation and choose Set verified GPU cap. Only a device-bound matching live cap is reported as verified.
4. Restart the same instance and run an identical benchmark. Once both runs are finished, in Power and hardware enter Baseline run ID and Capped run ID under Compare completed benchmark runs, then select Compare measured runs. The comparison requires matching GPU devices, instance, model name, depth, token count and repetitions, and clean reset-checked quick runs. It reports speed_cost_pct (positive means slower), mean watts and peak junction changes only when those sensors recorded measurements. Different hardware or workload settings are rejected; a cap is not automatically the sole cause of the change. Stop the instance again after measuring.
5. Enter the same PCI address, confirm and select Restore exact previous cap. This restores the saved exact microW value only if the same physical device and expected live cap are still present; otherwise inspect the mismatch and do not force a write. To leave the cap configured, skip restore intentionally and retain the cap policy.

The comparator is read-only, reports missing sensor data as unavailable and does not modify a power cap. Records absent from the recent benchmark history must be located in GPU tuning and benchmarks rather than guessed.

The guide power.cap-and-compare records the input fields pci, watts and instance. Its set step has a registered restore undo action. For a plain policy removal without live restoration, the separate Clear saved GPU power cap (leave live cap unchanged) control does not roll back the hardware cap.

Guided task: switch to stable maximum performance

1. Read Power profiles and drift and the built-in Stable performance profile. It requests supported performance-governor and platform low-idle-transition options; it is not a promise of maximum GPU clocks or measured improvement.
2. Under Guided task: switch to stable maximum performance, tick the confirmation after reviewing extra heat, sustained and idle draw, fan noise and rollback, then select Apply stable performance.
3. The page captures affected current values before applying. The helper verifies each selected setting. If an apply partly fails, the saved undo record remains for recovery; do not assume the machine is in a uniform state.
4. Observe CPU/thermal samples and compare an unchanged workload. Select Undo performance change to restore the recorded prior values. Undo requires successful per-setting readback. Only one pending performance undo is allowed, and this action does not persist a boot profile.
5. To keep settings through reboot, use Use at boot in Power options; high-impact PCIe, storage, GPU, turbo and watchdog settings require explicit review and force consent. Remove boot power profile reverses persistence, not necessarily live settings.

Telemetry, logging and tuning

Live CPU, GPU, power and temperature are shown from /api/hwinfo/live, /api/gpus, /api/gpu, /api/gpu-tune and other read-only routes. Sampler sequence, missing sensor reason and benchmark history matter when comparing runs. The Thermal logs capability draws from the power-budget and benchmark history reports; these are historical readings, not a substitute for a current device sensor. Review Power budget and UPS to compare estimates against supply and UPS ratings, and Boot stability and annotations to distinguish manually interrupted boots from crash evidence.

For GPU tuning, Apply tuning settings, Save GPU tuning profile and Set GPU profile at boot may change clocks, voltages, power and fan behavior. Repeated workload benchmarks are necessary; higher clocks can also increase watts, temperatures, throttling, reset risk and instability. GPU tuning and firmware backup/check remain separate from verified GPU power cap operations. GPU firmware inspection does not flash firmware.

Control reference

The help markers beside each native control use these exact explanations. No command-line flags are assigned to these controls in the UX registry; the API routes below are the executable interface.

| Capture power profile (power.profile-capture) | Capture the current readable power settings under a new name; capturing does not apply or enable a boot policy. | POST /api/power/profile/capture | | Disable power boot profile (power.boot-unpersist) | Stop persisting the selected boot profile without claiming to undo settings already applied to the machine. | POST /api/power/unpersist | | Save selected GPU tuning profile (power.tune-profile-save) | Save the selected GPU clock, fan, voltage or cap values as a profile without changing live hardware. | POST /api/gpu-tune/profile/save | | Capture GPU tuning profile (power.tune-profile-capture) | Capture readable current GPU settings as a profile; unreadable settings are omitted rather than assumed. | POST /api/gpu-tune/profile/save | | Apply GPU tuning profile (power.tune-profile-apply) | Apply a saved GPU tuning profile live; inspect each device readback and monitor heat before persisting. | POST /api/gpu-tune/profile/apply | | Persist GPU tuning profile (power.tune-profile-boot) | Select a verified GPU tuning profile for startup; boot-time changes can affect stability and heat. | POST /api/gpu-tune/profile/boot | | Disable GPU tuning boot profile (power.tune-profile-unboot) | Stop applying a GPU tuning profile at boot; this does not undo live changes already made. | POST /api/gpu-tune/profile/unboot | | Delete GPU tuning profile (power.tune-profile-delete) | Delete a custom GPU tuning profile after reviewing whether it remains selected at boot. | POST /api/gpu-tune/profile/delete | | Start GPU benchmark (power.bench-start) | Measure the same running workload repeatedly for a trustworthy GPU speed and power comparison. | POST /api/gpu-tune/bench/start | | Try GPU setting and benchmark (power.bench-trial) | Temporarily apply selected GPU settings while measuring performance; verify rollback and handle a failed restoration. | POST /api/gpu-tune/bench/start | | Test saved GPU profile (power.profile-test) | Benchmark the selected GPU profile as a temporary trial; compare with baseline and inspect restoration evidence. | POST /api/gpu-tune/bench/start | | Stop GPU benchmark (power.bench-stop) | Request cancellation of a running GPU benchmark and verify that it has stopped before further tuning. | POST /api/gpu-tune/bench/stop | | Delete benchmark history entry (power.bench-delete) | Delete one finished GPU benchmark record; running workloads are not stopped by deleting history. | POST /api/gpu-tune/bench/delete | | Back up GPU firmware (power.vbios-backup) | Request a read-only firmware backup from the selected GPU; never use this action to flash hardware. | POST /api/gpu-tune/vbios/backup | | Check local GPU ROM (power.rom-check) | Compare a local ROM image with the selected GPU without writing the image to the device. | POST /api/gpu-tune/rom-check |

Related guides

Hardware explains device detection and qualification. Panel Benchmarks and Evals describes performance measurement, and Reference Guided Tasks lists workflow conventions.