A simple yet powerful stress and benchmark utility for NVIDIA GPUs. CUDABurner is designed to test the performance and stability of your graphics card by running various compute-intensive CUDA kernels.
============================= CUDABurner v1.0 ==============================
[Global Info]
- Test Mode : benchmark
- Target GPU : [0] NVIDIA GeForce RTX 5060 Ti
----------------------------------------------------------------------------
[GPU 0 Vitals]
- Temperature : 70°C
- Power Usage : 147 W / 180 W (81.7%)
- GPU Clock : 2745 MHz
- Memory Clock: 13801 MHz
- Utilization : GPU 100% | VRAM 34%
----------------------------------------------------------------------------
[Benchmark Results] (In Progress...)
Precision Mode Performance Engine / Notes
----------------------------------------------------------------------------
FP64 Dense 0.33 TFLOPS CUDA Core (Native)
FP32 Dense 15.10 TFLOPS CUDA Core (TC Disabled)
TF32 Dense 24.26 TFLOPS Tensor Core (Default FP32)
FP16 Dense 49.39 TFLOPS Tensor Core
FP16 Sparse 96.68 TFLOPS Sparse Tensor Core
BF16 Dense 46.90 TFLOPS Tensor Core
INT8 Dense 51.33 TOPS Tensor Core
INT8 Sparse 367.87 TOPS Sparse Tensor Core
----------------------------------------------------------------------------
- NVIDIA Driver
- CUDA Toolkit (I use 13.1)
- CMake (>= 3.18)
- A C++17 compliant compiler
# Clone the repository
git clone https://github.com/stlin256/CUDABurner.git
cd CUDABurner
mkdir build
cd build
cmake ..
make -jThe executable CUDABurner will be located in the build directory.
# Run with interactive menu
./CUDABurner
# Run Benchmark Test (all precisions)
./CUDABurner --mode benchmark
# Run Benchmark Test (specific precisions)
./CUDABurner --mode benchmark --precision FP32 FP16 TF32
# Run Stress Test
./CUDABurner --mode stress- FP8/FP4 Not Implemented: Benchmark tests for FP8 and FP4 precision are not yet implemented.
- Insufficient Stress: The current stress-test (
power_virus_kernel) might not be sufficient to push GPUs to their absolute power limit.
一个简洁而强大的 NVIDIA GPU 压力与算力基准测试工具。CUDABurner 旨在通过运行多种计算密集型的 CUDA 核心，来测试您显卡的性能与稳定性。
============================= CUDABurner v1.0 ==============================
[Global Info]
- Test Mode : benchmark
- Target GPU : [0] NVIDIA GeForce RTX 5060 Ti
----------------------------------------------------------------------------
[GPU 0 Vitals]
- Temperature : 70°C
- Power Usage : 147 W / 180 W (81.7%)
- GPU Clock : 2745 MHz
- Memory Clock: 13801 MHz
- Utilization : GPU 100% | VRAM 34%
----------------------------------------------------------------------------
[Benchmark Results] (In Progress...)
Precision Mode Performance Engine / Notes
----------------------------------------------------------------------------
FP64 Dense 0.33 TFLOPS CUDA Core (Native)
FP32 Dense 15.10 TFLOPS CUDA Core (TC Disabled)
TF32 Dense 24.26 TFLOPS Tensor Core (Default FP32)
FP16 Dense 49.39 TFLOPS Tensor Core
FP16 Sparse 96.68 TFLOPS Sparse Tensor Core
BF16 Dense 46.90 TFLOPS Tensor Core
INT8 Dense 51.33 TOPS Tensor Core
INT8 Sparse 367.87 TOPS Sparse Tensor Core
----------------------------------------------------------------------------
- NVIDIA 驱动
- CUDA Toolkit (作者使用13.1)
- CMake (>= 3.18)
- 支持 C++17 的编译器
# 克隆仓库
git clone https://github.com/stlin256/CUDABurner.git
cd CUDABurner
mkdir build
cd build
cmake ..
make -j可执行文件 CUDABurner 将会生成在 build 目录下。
# 以交互式菜单运行
./CUDABurner
# 运行算力基准测试 (所有精度)
./CUDABurner --mode benchmark
# 运行算力基准测试 (指定精度)
./CUDABurner --mode benchmark --precision FP32 FP16 TF32
# 运行压力测试 (烤机)
./CUDABurner --mode stress- FP8/FP4 未实现: 针对 FP8 和 FP4 精度的算力基准测试尚未实现。
- 烤机压力不足: 当前的压力测试程序不足以让现代的高功耗 GPU 达到其极限功耗。