Would be possible to fix the fake "throttle"?

[NIICKTCHUNS]

Describe the bug
This is a known bug that make says that the GPU is thermal throttling or power throttling. Would be possible to fix the fake throttling info with RDNA3 GPUs by updating the "libdrm-amdgpu-sys" just like LACT did in the recent 0.8.0 update?

List relevant hardware/software information

• Arch Linux
• MangoHud version v0.8.1
• GPU RX 7600 Challenger OC

To Reproduce
Steps to reproduce the behavior:

1. Enable to show throttling and throttling graph
2. Enable mangohud in any game

Expected behavior
Only show throttling when the GPU is really throttling.

Screenshots

Additional context
Apparently this is a workaround as they say in the LACT update "Update libdrm-amdgpu-sys with RDNA3 hotspot throttle workaround" but it would be nice if this get implemented in Mangohud as well.

[17314642]

wait for v0.8.2 #1690

[sunnyyangyangyang]

report another fake throttling case on nvidia 5090

[flightlessmango]

@sunnyyangyangyang I think this case is legitimate. As the GPU approaches the TDP it will power throttle to keep it inside the boundaries

[sunnyyangyangyang]

@sunnyyangyangyang I think this case is legitimate. As the GPU approaches the TDP it will power throttle to keep it inside the boundaries

but i observe throttle multiple times , and as shown here there is only once gpu apporach to tdp. In fact i observe the throttling as many times as that when gpu power>500w, not actual tdp of 600w

[sunnyyangyangyang]

@sunnyyangyangyang I think this case is legitimate. As the GPU approaches the TDP it will power throttle to keep it inside the boundaries

as you can see the screenshot the fps remains constant even during "throttling" which is why i am concerning as during a real throttling it always result in fps drop.

[sunnyyangyangyang]

@sunnyyangyangyang I think this case is legitimate. As the GPU approaches the TDP it will power throttle to keep it inside the boundaries

My GPU runs AI model at 600W for 1 hour at only 68°C without throttling @2.83ghz. How can it throttle at 510W in games? The thermal headroom is huge!

[sunnyyangyangyang]

@sunnyyangyangyang I think this case is legitimate. As the GPU approaches the TDP it will power throttle to keep it inside the boundaries

UPDATE: I monitored with nvidia-smi during MangoHud's reported 'throttling'. Result: nvidia-smi shows ZERO throttling events while MangoHud claims throttling at 510W. This definitively proves it's a MangoHud detection bug for RTX 5090, similar to the RDNA3 issue.

[sunnyyangyangyang]

@flightlessmango

[flightlessmango]

nvidia-smi is an instant snapshot, so brief throttling blips may be missed.
MangoHud samples every 25 ms and reports if any throttling bits were set in the last 500 ms, so it catches short, non-constant events (as your frametime graph shows).
We don’t infer or simulate throttling, we display exactly what the driver reports via NVML

This is different from the AMD case. That was a confirmed driver bug where the hotspot bit was misreported, and we worked around it by excluding that bit.
If there is an NVIDIA driver misreport here, it needs to be confirmed and fixed by the driver team

[sunnyyangyangyang]

@flightlessmango

I have comprehensive evidence that this is a false throttling detection issue with RTX 5090, supported by NVIDIA's official NVML documentation and precise technical analysis.

NVML Bitmask Analysis (from NVIDIA official docs):

MangoHud's power throttling detection (src/gpu_metrics_util.h) uses:

if ((indep_throttle_status & 0x000000000000008CLL) != 0)
    power.push_back(0.1);

Breaking down 0x8C (10001100 binary):

• 0x04: nvmlClocksEventReasonSwPowerCap - "The clocks have been optimized to ensure not to exceed currently set power limits" (Source: https://docs.nvidia.com/deploy/nvml-api/group__nvmlClocksEventReasons.html)
• 0x08: nvmlClocksThrottleReasonHwSlowdown - "Power draw is too high and Fast Trigger protection is reducing the clocks" (Source: https://docs.nvidia.com/deploy/nvml-api/group__nvmlClocksEventReasons.html)
• 0x80: nvmlClocksThrottleReasonHwPowerBrakeSlowdown - "External Power Brake Assertion being triggered (e.g. by the system power supply)" (Source: https://docs.nvidia.com/deploy/nvml-api/group__nvmlClocksEventReasons.html)

The Critical Issue:
RTX 5090 triggers 0x04 (nvmlClocksEventReasonSwPowerCap) at ~510W as preventive power management, not performance throttling. NVIDIA defines this as "clocks have been optimized to ensure not to exceed currently set power limits" - this is power management optimization, not throttling!

Definitive Evidence:

nvidia-smi -q -d PERFORMANCE
SW Power Capping                  : 1089721985 us  # Microsecond precision!
HW Thermal Slowdown               : 0 us
HW Power Braking                  : 0 us

Technical Facts:

1. nvidia-smi precision: Tracks events with microsecond precision, 1000x more accurate than MangoHud's 25ms sampling
2. Zero performance impact: FPS remains constant during "throttling" - real throttling always causes performance drops
3. Power management ≠ Throttling: SW Power Capping is preventive optimization, not performance limitation
4. RTX 5090 power spikes: GPU has extreme transient spikes (900W+ for <1ms), triggering preventive management at 510W

Comparison with nvidia-smi during MangoHud's "throttling":

• MangoHud: Reports "power throttling"
• nvidia-smi: Shows "Not Active" for all throttling states
• Result: Zero actual throttling events detected by official NVIDIA tool

Conclusion:
This is identical to the AMD RDNA3 issue - MangoHud misinterprets preventive power management flags as performance throttling. RTX 5090 needs a similar workaround to exclude the 0x04 bit (nvmlClocksEventReasonSwPowerCap) from power throttling detection.

Proposed Fix:
Modify RTX 5090 detection to exclude preventive power management bits, similar to the RDNA3 hotspot workaround.

[flightlessmango]

This is identical to the AMD RDNA3 issue

RDNA3 issue is a bug in the driver, this is working as intended.

preventive power management, not performance throttling

Power management works by throttling to stay within the power limit. If the word throttling is the concern, that is NVIDIA’s own terminology in NVML. MangoHud reflects those NVML events as reported. AMD exposes equivalent power cap flags and we display those as well.

[guskikalola]

@sunnyyangyangyang What monitoring software are you using besides MangoHud?

[sunnyyangyangyang]

@sunnyyangyangyang What monitoring software are you using besides MangoHud?

nvidia-smi

[guskikalola]

@sunnyyangyangyang What monitoring software are you using besides MangoHud?

nvidia-smi

Is that graph you show from an nvidia smi gui?

[sunnyyangyangyang]

@sunnyyangyangyang What monitoring software are you using besides MangoHud?

nvidia-smi

Is that graph you show from an nvidia smi gui?
https://github.com/codifryed/coolercontrol