GPU-PCIe-Test

A tool to benchmark GPU/PCIe bandwidth, latency, and VRAM integrity. Measures real-world data transfer speeds between CPU and GPU memory.

Features

Bandwidth Testing

• Upload Speed - CPU to GPU transfer rates
• Download Speed - GPU to CPU transfer rates
• Bidirectional - Simultaneous upload/download
• Latency - Round-trip timing measurements

VRAM Integrity Scanning

• Multi-chunk testing - Tests up to 80-90% of VRAM by cycling through multiple allocations
• 8 test patterns - Zeros, ones, checkerboard, inverse checkerboard, address, random, marching ones, marching zeros
• Error clustering - Groups nearby faults for easier diagnosis
• Progress tracking - Real-time progress with cancellation support

Hardware Detection

• PCIe generation & lanes - Detects PCIe 3.0/4.0/5.0 and x1-x16 configurations
• Thunderbolt/USB4 - Identifies external GPU connections
• OCuLink - Detects OCuLink eGPU setups
• Integrated GPU detection - Identifies iGPUs and shows system RAM bandwidth comparison
• System RAM info - Displays DDR4/DDR5/LPDDR5 speed, channels, and theoretical bandwidth

Results & Export

• Min/Avg/Max graphs - Visual bandwidth comparison
• Ranked comparison - Compare results against PCIe/TB/USB4 standards
• CSV export - Export benchmark data for analysis
• Clipboard copy - Quick copy of VRAM scan results

Screenshots

Coming soon

Requirements

Windows

• Windows 10/11 (64-bit)
• DirectX 12 compatible GPU
• Visual Studio 2019+ (for building)

Linux

• Vulkan-capable GPU with drivers installed
• CMake 3.16+, g++, libglfw3-dev, libvulkan-dev
• X11 or Wayland display server

Installing (prebuilt)

Download from the Releases page.

Platform	File	Notes
Windows (Direct3D 12)	GPU-PCIe-Test_GUI.exe	Primary build; no installer, just run it
Windows (Vulkan)	GPU-PCIe-Test-Vulkan.exe	Needs a Vulkan-capable driver
Linux (any distro)	GPU-PCIe-Test-<version>-x86_64.AppImage	chmod +x and run; native Wayland or X11 (bundled GLFW 3.4); uses your distro's Vulkan loader and GPU driver
Arch Linux / CachyOS	gpu-pcie-test-<version>-1-x86_64.pkg.tar.zst	sudo pacman -U <file>, or build it yourself: cd packaging/arch && makepkg -si

Run the Linux build as root (sudo) if you want full system RAM details from dmidecode.

Building

Windows - Quick Build (Recommended)

1. Clone the repository:

   git clone https://github.com/djanice1980/GPU-PCIe-Test.git
   cd GPU-PCIe-Test

2. Run the build script:

   build_gui.bat

   The script automatically downloads ImGui and ImPlot dependencies.

3. Run the executable:

   GPU-PCIe-Test.exe

Windows - Manual Build

If you prefer manual compilation:

cl /EHsc /O2 /DUNICODE /D_UNICODE main_gui.cpp imgui*.cpp implot*.cpp ^
   /link d3d12.lib dxgi.lib d3dcompiler.lib user32.lib gdi32.lib ^
   /out:GPU-PCIe-Test.exe

Linux Build

cd Linux
chmod +x build_linux.sh
./build_linux.sh
./build/gpu-pcie-test-vulkan

See Linux/Linux_README.md for detailed instructions and distro-specific package names.

Usage

Bandwidth Benchmark

1. Select your GPU from the dropdown
2. Adjust test parameters (buffer size, iterations) if desired
3. Click Run Benchmark
4. View results in the graph and log panels

VRAM Scan

1. Select a discrete GPU (not available for integrated GPUs)
2. Pick a Scan preset (or use Custom for fine control):
   • Quick - 4 patterns, 50% coverage (~30 sec sanity check)
   • Standard - 8 patterns, 80% coverage (~2-3 min, default)
   • Deep - 8 patterns + refresh check + address-bus check, 90% coverage (~5 min)
   • Thorough - same as Deep with 10x re-reads, 95% coverage (~12-15 min)
3. Optionally fine-tune individual options:
   • Patterns - enable/disable any of 8 test patterns
   • Refresh check - re-read each chunk multiple times to catch retention errors
   • Address-bus check - randomize block read order to expose address-bus errors
   • Coverage - choose 50/80/90/95% of VRAM
4. Click VRAM Scan
5. Wait for completion - test covers multiple memory regions
6. Results window includes a bit-level Error Breakdown classifying any errors found (single-bit, multi-bit, address-bus, stuck-bit, refresh)

Note: VRAM scanning is a basic integrity test. For chip-level diagnosis, use vendor tools like NVIDIA MATS or AMD memory diagnostics. For extended overclock stress testing, consider memtest_vulkan.

How It Works

Bandwidth Testing

Uses D3D12 copy operations between upload heaps (CPU-visible) and default heaps (GPU-local) with fence synchronization for accurate timing.

VRAM Scanning

1. Allocates 512MB test buffers (upload, GPU, readback)
2. For each chunk (up to 25+ chunks for 80% coverage):
   • Writes test pattern to upload buffer
   • Copies to GPU buffer
   • Copies back to readback buffer
   • Compares against expected pattern
   • Releases and reallocates to potentially hit different physical VRAM
3. Reports errors by pattern type and memory region

Interface Detection

Queries PCIe link status registers via DXGI and applies heuristics for:

• Thunderbolt (x4 links with specific bandwidth signatures)
• USB4 (similar to TB but different enumeration patterns)
• OCuLink (x4/x8 with PCIe-level bandwidth)

Limitations

• D3D12 abstraction - Cannot directly address physical VRAM; relies on driver allocation patterns
• No stress testing - Tests static memory, not thermal/power stress conditions
• D3D12 variant is Windows only - The Vulkan variant runs on both Windows and Linux
• No multi-GPU - Tests one GPU at a time

Version History

See CHANGELOG.md for detailed release notes.

Contributing

Contributions welcome! Please feel free to submit issues and pull requests.

License

MIT License - see LICENSE for details.

Author

David Janice
Email: djanice1980@gmail.com
GitHub: @djanice1980

Acknowledgments

• Dear ImGui - Immediate mode GUI
• ImPlot - Plotting library for ImGui
• Microsoft DirectX 12 documentation