WTG: NVML memory-utilization counter pinned at 100% idle on tested RTX 3080 Laptop GPU drivers

adam.hooper · May 10, 2026, 11:10pm

Posting here because the System Management and Monitoring (NVML) category appears read-only for new topics. This topic is about NVML/NVIDIA-SMI behavior, not NVAPI.

I built a Windows-native NVML telemetry tool called WTG and found a reproducible telemetry anomaly while regression-testing it across several NVIDIA GPU/driver combinations.

WTG is intended to be a small reference instrument for asking a direct question: what does the NVIDIA driver report through NVML when queried for GPU utilization, memory-controller activity, VRAM allocation, power, process attribution, and runtime context?

Repository:

Regression research note:

github.com/novovictus/WTG

artifacts/regression-research/wtg_regression_research.md

main

# WTG Regression Research: A Mobile Ampere Memory-Utilization Counter Goes to 100 Percent at Idle

Windows GPU telemetry is layered. Task Manager can show that a GPU is busy, and vendor tools can expose selected counters, but the CUDA-relevant truth layer lives closer to the NVIDIA driver. WTG was built to operate at that layer. It is a Windows-native NVML reference tool intended to answer a direct question: what does the NVIDIA driver report when queried for GPU utilization, memory-controller activity, VRAM allocation, power, process attribution, and runtime context?

During WTG regression testing, one signal separated itself from normal variance. On a consumer mobile Ampere platform, the NVIDIA GeForce RTX 3080 Laptop GPU began reporting 100 percent memory-controller utilization at idle or near-idle beginning with the 580.88 driver branch. The rest of the telemetry did not agree with real saturation: GPU utilization stayed near zero, VRAM allocation stayed low, thermals stayed normal, and power remained in an idle or light-load range.

This report is not a bug report against WTG. WTG is the reference instrument used to expose and compare the driver behavior. The finding is a driver-facing telemetry regression candidate: a reproducible NVML memory-utilization anomaly on a specific consumer mobile GPU class and driver branch, bounded by earlier-driver controls, mobile negative controls, and a stable desktop bench.

## What the counter means

The key field is NVML memory utilization, represented in WTG's beta 4 probe output as `util.mem_controller_pct`. This field is not VRAM occupancy. It is a utilization counter for memory read/write activity during the sample interval. VRAM occupancy is captured separately as `vram.used_mib` and `vram.total_mib`.

That distinction is essential. A GPU can have low VRAM allocation while still performing memory reads and writes, and it can also have VRAM allocated while doing little active work. The anomaly here is not that the GPU has allocated memory. The anomaly is that the memory-controller activity counter is pinned at 100 percent when the rest of the telemetry indicates idle or near-idle behavior.

## The regression signal

The affected system class is the ASUS ROG Strix G533QS_G533QS with the NVIDIA GeForce RTX 3080 Laptop GPU under Windows WDDM. The RTX 3080 Laptop GPU evidence was not collected from a single machine only; the regression set includes samples from two identical ASUS G533QS systems with the same GPU class. Earlier drivers form the baseline. With driver 566.07, the RTX 3080 Laptop GPU reported 0 percent GPU utilization and 0 percent memory-controller utilization, with low VRAM allocation and normal power. With driver 577.00, the same ASUS RTX 3080 Laptop GPU platform again reported normal memory-controller behavior: low GPU activity, 0 percent memory-controller utilization, modest VRAM allocation, and normal power.

The behavior changes at driver 580.88. In the baseline capture, the RTX 3080 Laptop GPU reported 0 percent GPU utilization and 100 percent memory-controller utilization while using only 292 MiB of 16,384 MiB of VRAM. Power was approximately 17.5 W against a 130 W limit, and the parallel NVIDIA-SMI snapshot showed 0 percent GPU utilization and only 86 MiB of 16,384 MiB frame-buffer memory in use. That is not the shape of a saturated memory workload.

This file has been truncated. show original

Test matrix:

github.com/novovictus/WTG

artifacts/test-matrix/matrix.md

main

# WTG Test Matrix (Early Validation)

WTG version tested baseline: v0.1.2+ (truth-layer hardened)

This matrix records **empirical test results** collected during early WTG development. It documents observed behavior only and does not imply official support or compatibility guarantees.

Raw run logs are not published to avoid exposing host-specific identifiers.

---

## Tested GPU / Driver Combinations

| GPU Model                 | Architecture | Driver Version(s)        | CUDA Version(s)    | OS        | NVML Load   | NVML mem util (Idle)                | --once   | --watch   | Notes                                                                                                    |
| ------------------------- | ------------ | ------------------------ | ------------------ | --------- | ----------- | ----------------------------------- | -------- | --------- | -------------------------------------------------------------------------------------------------------- |
| RTX 3080 Laptop GPU       | Ampere       | 566.07                   | 12.7               | Win11 x64 | Yes         | OK                                  | Yes      | Yes       | Stable                                                                                                   |
| RTX 3080 Laptop GPU       | Ampere       | 577.00                   | 12.9               | Win11 x64 | Yes         | OK                                  | Yes      | Yes       | Stable                                                                                                   |
| RTX 3080 Laptop GPU       | Ampere       | 580.88+ (conf. to 591.86)| 13.0+              | Win11 x64 | Yes         | 100% @ idle (first observed 580.88) | Yes      | Yes       | Driver-branch regression; normal thermals/power; WDDM consumer mobile SKU                                |
| RTX 4070 Laptop GPU       | Ada          | 591.59                   | 13.1               | Win11 x64 | Yes         | OK                                  | Yes      | Yes       | Stable; transient `nvidia-smi` formatting anomaly observed                                               |
| RTX 3060 Laptop GPU       | Ampere       | 512.74 / 580.88 / 591.86 | 11.6 / 13.0 / 13.1 | Win11 x64 | Yes         | OK                                  | Yes      | Yes       | 512: cap unavailable; 580: 105W cap exposed; 591: 80W cap (policy shift); no mem-util regression         |
| RTX 3060 Ti (Desktop OEM) | Ampere       | 580.88 / 581.04          | 13.0               | Win11 x64 | Yes         | OK                                  | Yes      | Yes       | Desktop reference; 200W cap exposed; idle + sustained load validated (≈199W / 200W); no mem-util anomaly |

This file has been truncated. show original

Bench baseline:

github.com/novovictus/WTG

artifacts/regression-research/bench_v1.0.0.md

main

# WTG Bench v1 Hardware and Software Baseline

## Status

Active mixed adapter WDDM configuration Intel iGPU is primary display
RTX 3060 Ti is installed as headless compute device

Firmware validated via GPU-Z: Above 4G Decoding: Enabled (firmware
managed, not user configurable) Resizable BAR: Not supported by current
GPU / VBIOS / driver combination BAR1 Aperture: 256 MiB (legacy mode)

------------------------------------------------------------------------

## Platform

### System Lineage

Dell Inspiron 3891 Service Tag: 6N2R0M3 Open air 2020 T rail style bench
ATX PSU conversion

This file has been truncated. show original

The narrow finding:

On tested ASUS ROG Strix G533QS systems with GeForce RTX 3080 Laptop GPU under Windows WDDM, NVML memory utilization begins reporting 100% at idle or near-idle starting with the 580.88 driver branch. Adjacent telemetry does not support real memory saturation: GPU utilization remains near zero, VRAM allocation remains low, thermals remain normal, power remains in the idle/light-load range, and NVIDIA-SMI comparison output also shows low activity.

This is not a claim that applications are affected. It is not a claim of VRAM exhaustion. It is not a claim that WTG has identified the internal NVIDIA driver cause. It is a driver-facing telemetry inconsistency candidate exposed while validating a tool that queries NVML directly.

Current evidence summary:

Affected sample set:

ASUS ROG Strix G533QS_G533QS
GeForce RTX 3080 Laptop GPU
Windows WDDM
Two identical ASUS G533QS systems represented in the regression set

Earlier-driver baseline:

566.07: normal idle memory-controller utilization
577.00: normal idle memory-controller utilization

Later-driver behavior:

580.88: memory-controller utilization reports 100% at idle/near-idle
591.59: same pattern persists

Controls:

Desktop RTX 3060 Ti bench reports coherent idle and load telemetry
Acer Nitro AN517-42 with RTX 3060 Laptop GPU on 580.88 does not reproduce the idle 100% memory-utilization symptom

The key distinction is that this is NVML memory utilization, not VRAM allocation. In WTG beta 4 probe output, the field is labeled util.mem_controller_pct to avoid confusing memory-controller activity with vram.used_mib.

The desktop control is important because it shows WTG reporting expected behavior under real load: high GPU utilization, high memory-controller utilization, high power, elevated performance state, and substantial VRAM allocation all move together. On the affected mobile RTX 3080 Laptop GPU captures, the memory-controller counter is pinned while the surrounding telemetry remains in an idle or light-load envelope.

What I am looking for:

Confirmation whether this behavior is expected for this class of mobile GeForce/WDDM platform.
Comparable validation from other RTX 3080 Laptop GPU systems, especially ASUS mobile Ampere systems.
Any known driver-branch changes around 580.88+ that could explain this NVML memory-utilization counter behavior.
Suggestions for additional NVML fields or NVIDIA-SMI queries that would better separate driver-reporting behavior from actual workload activity.

I am intentionally not generalizing this to all mobile Ampere GPUs or all 580.88+ drivers. The current data supports a narrower finding: the tested ASUS RTX 3080 Laptop GPU platforms show a 100% idle/near-idle NVML memory-utilization counter on later driver branches while surrounding telemetry remains inconsistent with real saturation.

Topic		Replies	Views
RTX 2080 Ti always Power Cap and low utilization Linux	14	5192	November 6, 2019
Overclocking doesn't work on Maxwell GPUs System Management and Monitoring (NVML)	21	1210	October 21, 2026
How to get gpu Utilization with nvmlDeviceGetUtilizationRates() System Management and Monitoring (NVML)	2	1980	July 18, 2019
Increased idle consumption with driver 570 Linux	25	5594	June 23, 2025
Nvidia-smi and nvmlDeviceGetUtilizationRates do not match System Management and Monitoring (NVML)	0	1194	May 24, 2022
Nvidia driver on Ubuntu 24.04.04 LTS with AMD threadripper 3995WX + Nvidia RTX 3080 Ti fails and gdm-x-session deletes GPU-0 Linux	5	547	March 21, 2026
Nvidia-smi show H100 run at full load but power consumption only ~110w Drivers - Linux, Windows, MacOS	1	1419	March 16, 2025
[Bug] Xid 62 GSP Watchdog Timeout on RTX 5080 during high-context LLM inference Linux driver , linux-driver , rtx , llama	5	614	July 11, 2026
On Windows, getting list of graphics or compute processes can return invalid argument depending on launch System Management and Monitoring (NVML)	32	2976	July 6, 2024
`nvmlDeviceGetComputeRunningProcesses_v3()` inside container reporting invalid memory usage number System Management and Monitoring (NVML)	0	464	May 30, 2024

WTG: NVML memory-utilization counter pinned at 100% idle on tested RTX 3080 Laptop GPU drivers

Related topics