An advanced GPU hardware diagnostic, multi-pattern VRAM integrity tester, and thermal telemetry monitor for NVIDIA RTX 4090 and modern GeForce GPUs (Ada Lovelace, Ampere, and Blackwell architectures).
Built with PyTorch CUDA, PyQt6, and native NVAPI / MMIO register telemetry.
- Real-Time VRAM GDDR6X / GDDR6 / GDDR7 Memory Temperature Telemetry:
- Live junction temperature monitoring directly from hardware sensors (channel 7 on RTX 4090).
- Real-time GPU Hot Spot and GPU Core temperatures.
- Multi-line rolling 60-second waveform graph with dynamic thermal alert thresholds.
- Advanced Multi-Pattern VRAM Memory Error & Bit-Retention Test:
- Solid Zeros (
0x00000000) & Solid Ones (0xFFFFFFFF): Stuck-at-0 and stuck-at-1 physical cell diagnostics. - Checkerboard (
0x55555555/0xAAAAAAAA): Adjacent capacitor crosstalk, line interference, and signal coupling checks. - Walking Inversions (Walking 1s): Memory address line and trace integrity validation.
- High-Entropy Random Noise: Long-hold bit-flip detection under sustained VRAM bus saturation.
- Precise bit-flip diagnostic reporting: chunk index, byte offset, expected vs received hex values, and binary XOR bit-flip masks.
- Solid Zeros (
- Tensor Core Arithmetic Precision & Stability Verification:
- Full matrix multiplication (GEMM) stress tests in FP64, FP32, FP16, and BF16.
- NaN/Inf detection, precision drift validation against CPU golden reference, and TFLOPS measurement.
- PCIe & Memory Bandwidth Benchmark:
- Measures Host-to-Device (H2D), Device-to-Host (D2H), and internal VRAM-to-VRAM copy bandwidth (GB/s).
- Modern Hardware-Themed PyQt6 GUI:
- Cyberpunk dark UI styled for NVIDIA RTX hardware.
- Thread-safe background worker keeping the UI smooth at 60 FPS under heavy compute and DMA stress.
- Drag-and-drop report directory box (accepts dropped folders from Windows Explorer).
- One-click timestamped diagnostic report export.
- Full CLI & Headless Support:
- Runs headless or automated via command-line arguments.
This project builds upon and credits the work of two open-source projects:
- Holozing1/vram-guard:
- Provided foundational methodology for VRAM stress testing, multi-pattern memory verification, and bit-flip integrity analysis.
- olealgoritme/gddr6:
- Provided the reverse-engineered hardware register offsets and decoding logic for reading GDDR6, GDDR6X, and GDDR7 memory temperatures directly from DRAM sensors.
- OS: Windows 10/11 (64-bit) or Linux
- GPU: NVIDIA RTX 40-series (RTX 4090, 4080, etc.), RTX 30-series (3090, 3080, etc.), or RTX 50-series
- Python: 3.10 or newer (Python 3.12 recommended)
- NVIDIA Driver: 522.25+ recommended
Simply download or clone the repository and double-click start.bat:
git clone https://github.com/zabique/test_vram.git
cd test_vram
start.batstart.bat automatically:
- Detects
uv(fast package manager) or standardpython. - Creates the
.venvvirtual environment. - Automatically downloads and installs PyTorch with CUDA 12.6 and PyQt6.
- Launches the desktop application.
If you have uv installed:
# 1. Clone repository
git clone https://github.com/zabique/test_vram.git
cd test_vram
# 2. Create virtual environment
uv venv --python 3.12 .venv
# 3. Activate virtual environment
# On Windows:
.venv\Scripts\activate
# On Linux:
source .venv/bin/activate
# 4. Install dependencies
uv pip install torch --index-url https://download.pytorch.org/whl/cu126 PyQt6
# 5. Launch
python main.pyWorks on any Windows or Linux machine with standard Python:
# 1. Clone repository
git clone https://github.com/zabique/test_vram.git
cd test_vram
# 2. Create virtual environment
python -m venv .venv
# 3. Activate virtual environment
# Windows:
.venv\Scripts\activate
# Linux:
source .venv/bin/activate
# 4. Install dependencies from requirements.txt
pip install -r requirements.txt
# 5. Launch
python main.pypython main.py(or double-click start.bat)
Query current GPU core, hotspot, and GDDR6X VRAM junction temperatures once:
python main.py --read-tempSample output:
=================================================================
NVIDIA GPU & VRAM THERMAL STATUS
=================================================================
Device Name : NVIDIA GeForce RTX 4090
Telemetry Engine : NVAPI (GetThermalSensors)
GPU Core Temp : 46.2 °C
GPU Hot Spot : 55.1 °C
VRAM Memory Temp : 52.0 °C (GDDR6X/GDDR6 Junction)
Continuously stream live temperatures in your terminal:
python main.py --monitor-temp --temp-interval 1.0Run the full diagnostic suite headlessly with custom allocation size and test options:
# Test 16 GiB with all patterns
python main.py --cli --gib 16 --rounds 2
# Quick 4 GiB test skipping bandwidth & tensor tests
python main.py --cli --gib 4 --quick --skip-bandwidth --skip-tensorgui.py: PyQt6 desktop application with real-time waveform graph, telemetry cards, test controls, and drag-and-drop report export.main.py: Main entrypoint with dual-mode dispatcher (GUI by default, CLI on argument flag).gpu_thermal.py: Unified thermal telemetry engine supporting Windows native NVAPI (0x65FE3AAD), Linux direct BAR0 MMIO (olealgoritme/gddr6), and NVML fallbacks.start.bat: One-click launcher and automateduvvirtual environment setup script.
MIT License. See LICENSE for details.