NVCheckup

Cross-platform NVIDIA diagnostics for gaming, AI, and creators.

Read-only by default · Privacy-first · Open Source

Download View on GitHub Build from source
v0.2.3
Windows + Linux
x86_64 + ARM64
DGX Spark + RTX Spark · simulated
MIT License
nvcheckup — full diagnostic
$ nvcheckup run --mode full --zip

[1/7] Collecting system information...
[2/7] Detecting GPUs and drivers...
[3/7] Collecting GPU thermal and PCIe data...
[4/7] Running platform-specific checks...
[5/7] Checking AI/CUDA environment...
[6/7] Skipping network probes (use --network to enable)...
[7/7] Analyzing results...

  Written: report.txt
  Bundle:  nvcheckup-bundle-20260901-143210.zip

NVCheckup v0.2.3 | 2026-09-01 14:32:10
OS: Microsoft Windows 11 Pro 10.0.26100 | Arch: amd64
GPU: NVIDIA GeForce RTX 4070 | Driver: 591.86 | VRAM: 12282 MB
CUDA (driver): 13.1 | CUDA Toolkit: 12.8
PyTorch: 2.8.0+cpu (CUDA NOT available)
Temp: 86°C | P-State: P0 | Util: 98%
PCIe: Gen4 x16
Findings: 1 CRITICAL, 2 WARNING, 5 total | 1 auto-fixable
Top: Display Driver Resets Detected (Event ID 4101); PyTorch Installe...

Top Issues:
  1. [CRIT] Display Driver Resets Detected (Event ID 4101) (92% confidence)
  2. [WARN] PyTorch Installed Without CUDA Support (95% confidence)
  3. [WARN] GPU Running Hot (80% confidence)

What NVCheckup Detects

🎮

Driver Crashes

Scans Windows Event Logs for Event ID 4101, nvlddmkm errors, and WHEA hardware faults. Correlates with recent Windows Updates.

🧠

CUDA Stack Issues

Detects driver/toolkit version mismatches, missing libcuda.so, broken DKMS builds, and Secure Boot blocking NVIDIA kernel modules.

🐍

PyTorch + TensorFlow

Probes torch.cuda.is_available(), detects CPU-only wheels, version mismatches, and environment confusion between conda/venv.

💻

Hybrid GPU

Detects NVIDIA + iGPU configurations, PRIME offloading state, and applications potentially running on the wrong GPU.

🎥

Overlay Conflicts

Identifies Xbox Game Bar, Discord, RTSS, and other overlay software that may contribute to stutter or frame pacing issues.

⚡
New · simulated, unverified on hardware

DGX Spark & RTX Spark

Recognises GB10 and N1X unified-memory platforms, reads the shared pool from /proc/meminfo instead of the [N/A] that nvidia-smi reports by design, silences the VRAM, fan and PCIe false alarms, checks DGX OS driver/firmware pairing and the ConnectX-7 twins, and sizes LLM deployments with nvcheckup llm-plan. Read the section · own one? issue #3

📦

Support Bundles

Generates redacted reports in TXT, JSON, and Markdown. Packages everything into a zip bundle ready for GitHub issues or forums.

Built for Trust

✔ Read-only by default; opt-in fixes are confirmed and undoable
✔ PII redaction ON
✔ No network calls unless you opt in with --network
✔ No background services
✘ No telemetry
✘ No driver modification
✘ No silent registry edits
✘ No data uploads
✔ Spark fixes are printed as Advisory steps with a revert command, never applied