GPU-Burn

A multi-GPU CUDA stress testing tool with monitoring and multiple modes.

Features

• Multiple Modes: GEMM (matrix multiply), Compute (ALU+SFU), Memory (bandwidth + random access)
• GPU Monitoring: Temperature, power, clocks, and throttling detection via NVML
• Multi-GPU Support: Test all GPUs simultaneously with independent error tracking
• Configurable Parameters: Matrix sizes, memory usage, precision (FP32/FP64), tensor cores
• Cross-Platform: Windows and Linux support

Building

Prerequisites

• CMake 3.24+
• CUDA Toolkit 11.0+
• NVML library (included with NVIDIA drivers)

Build Instructions

mkdir build && cd build
cmake ..
cmake --build . --config Release

Prebuilt Binaries

Download from Releases. Binaries are dynamically linked and require matching CUDA installed, check your version with nvcc --version or nvidia-smi.

Binary	Requires
gpu-burn-linux-cuda13.1.0.tar.gz	CUDA 13.1
gpu-burn-linux-cuda12.8.1.tar.gz	CUDA 12.8
gpu-burn-linux-cuda11.8.0.tar.gz	CUDA 11.8
gpu-burn-windows-cuda13.1.0.exe	CUDA 13.1
gpu-burn-windows-cuda12.8.1.exe	CUDA 12.8

Usage

# Basic usage - test all GPUs for 60 seconds
gpu-burn 60

# Use 90% of GPU memory instead of default 80%
gpu-burn -m 90% 300

# Test specific GPU
gpu-burn -i 0 120

# Verbose monitoring with detailed metrics
gpu-burn -v 60

# Different stress modes
gpu-burn -mode compute 120     # ALU+SFU compute stress
gpu-burn -mode memory 120      # Memory bandwidth stress
gpu-burn -mode gemm 120        # Matrix multiply stress (default)

# Larger matrices for maximum stress
gpu-burn -s 16384 300

# Double precision (FP64)
gpu-burn -d 300

# Try tensor cores (if supported)
gpu-burn -tc 300

Command-Line Options

Option	Description
-m X	Use X MB of GPU memory
-m N%	Use N% of available GPU memory (default: 80%)
-s N	Matrix size NxN for GEMM mode (default: 8192, must be multiple of 256)
-mode M	Stress mode: gemm, compute, memory (default: gemm)
-d	Use double precision (FP64)
-tc	Use tensor cores (only in GEMM mode; uses TF32 for FP32, requires Ampere+)
-l	List GPUs and exit
-i N	Test GPU index N only (0-based)
-v	Verbose output
-h	Show help

Stress Modes

GEMM Mode (default)

Matrix multiplication stress using cuBLAS. -s option controls matrix size.

Compute Mode

ALU and SFU (Special Function Unit) stress with minimal memory bandwidth requirements.

Memory Mode

Memory stress testing with four patterns that cycle continuously (not well designed currently):

1. Sequential Access (with validation): Coalesced memory reads/writes
2. Random Access (read-only): Random memory access
3. Stride Access (read-only): Strided access to cause thrashing
4. PCIe Transfer (with validation): Host-Device transfers

Monitoring Output

Normal Mode

Use the same output format from the original gpu-burn:

89.2%  proc'd: 1234 (7823.5 Gflop/s)  errors: 0  72C 245W

Shows: Progress %, iterations processed, performance, errors, temperature, power, throttling indicator (!)

Verbose Mode

89.2%  proc'd: 1234 (7823.5 Gflop/s)  errors: 0
GPU 0: 72C  245W  SM:1890MHz  Mem:5001MHz

Shows: Detailed per-GPU metrics including clock speeds and throttling reasons

Error Detection

All modes include built-in validation (not exhaustive):

• GEMM mode: Compares matrix multiplication results against reference
• Compute mode: Validates ALU and SFU computation consistency
• Memory mode: Validates sequential and PCIe patterns every 4th iteration

Credits

Based on the original gpu-burn, rewritten with:

• NVML integration for real-time monitoring
• Multiple stress modes (GEMM, Compute, Memory)
• Cross-platform support (Windows/Linux)