RTX 5000 Ada Laptop GPU reports invalid 588–593 W power draw while Runtime D3 suspended

Graphics / Linux Linux
, , ,
1

Hello,

I am seeing invalid NVML / nvidia-smi power telemetry on an NVIDIA RTX 5000 Ada Generation Laptop GPU in a Dell Precision 7780.

System

  • System: Dell Precision 7780
  • GPU: NVIDIA RTX 5000 Ada Generation Laptop GPU
  • PCI ID: 10de:2730
  • Subsystem ID: 1028:0c42
  • VBIOS: 95.03.2A.00.49
  • System BIOS: Dell 1.29.0
  • Distribution: Debian 13
  • Kernel: 6.12.100+deb13-amd64
  • NVIDIA driver: 550.163.01
  • CUDA version reported by nvidia-smi: 12.4

The installed driver is the Debian-packaged NVIDIA driver, not a Dell-supplied graphics driver.

Problem

When the GPU is in PCI Runtime D3 suspend, Linux reports:

$ cat /sys/bus/pci/devices/0000:01:00.0/power/runtime_status
suspended

However, NVML / nvidia-smi reports physically impossible power values:

$ nvidia-smi --query-gpu=persistence_mode,power.draw,pstate,utilization.gpu --format=csv

persistence_mode, power.draw [W], pstate, utilization.gpu [%]
Disabled, 593.51 W, P0, 0 %

Other observed invalid values include approximately:

588.21 W
593.51 W

The reported power limits are:

Current Power Limit: 115.00 W
Default Power Limit: 115.00 W
Minimum Power Limit:   5.00 W
Maximum Power Limit: 145.00 W

The reported power draw is therefore far above the GPU’s configured maximum.

State during the invalid reading

  • GPU utilization: 0 %
  • memory utilization: 0 %
  • no CUDA or graphics processes
  • no process has /dev/nvidia* open
  • Display Active: Disabled
  • Display Mode: Disabled
  • PCI runtime status: suspended
  • no NVRM Xid errors
  • no PCIe errors

The clocks and P-state also appear stale or invalid while suspended:

Performance State: P0
Graphics Clock: 1920 MHz
Memory Clock: 9000 MHz
Power Draw: 593.51 W

When the GPU is temporarily awake, plausible values such as approximately 9–37 W are reported. After it returns to Runtime D3 suspend, the invalid 588–593 W readings appear again.

Enabling persistence mode briefly produced plausible values, but the invalid readings returned later, so persistence mode is not a reliable workaround.

Reproduction

cat /sys/bus/pci/devices/0000:01:00.0/power/runtime_status

nvidia-smi \
  --query-gpu=power.draw,power.draw.instant,pstate,utilization.gpu \
  --format=csv

Typical result:

suspended

power.draw [W], power.draw.instant [W], pstate, utilization.gpu [%]
593.51 W, 593.51 W, P0, 0 %

Expected behavior

nvidia-bug-report_magister-primus_2026-08-01.log.gz (2.1 MB)

If power telemetry is unavailable while the GPU is in Runtime D3 suspend, NVML / nvidia-smi should return something such as:

N/A
Suspended
Not Available

It should not return an apparently valid numeric value above the GPU’s maximum power limit.

Possible improvements would be:

  • return N/A for unavailable telemetry during Runtime D3
  • mark power, clock, and P-state values as stale while suspended
  • expose Runtime D3 state through NVML
  • reject power values exceeding the reported maximum power limit

The current behavior can cause false monitoring alarms, invalid Prometheus or Grafana metrics, incorrect energy calculations, and confusion during diagnostics.

I understand that Runtime D3 behavior may also depend on OEM BIOS, ACPI, embedded-controller, or GPU-VBIOS implementation. If this is caused by Dell platform firmware rather than NVML, please advise which component should be reported to Dell.

I have generated an nvidia-bug-report.log.gz and can attach it if required.