vkpeak

A synthetic benchmarking tool to measure peak capabilities of vulkan devices. It only measures the peak metrics that can be achieved using vector operations and does not represent a real-world use case.

Download

Download Windows/Linux/MacOS Executable for Intel/AMD/Nvidia/Apple GPU

https://github.com/nihui/vkpeak/releases

Usages

vkpeak.exe

vkpeak will choose the default vulkan device (device id 0).

If you need to specify device id, then

vkpeak.exe 0

The device_id parameter is optional and defaults to 0.

By default, if you do not pass a scenario argument, vkpeak runs all available scenarios.

You can optionally select one scenario:

vkpeak.exe fp16-matrix

Or with an explicit device id:

vkpeak.exe 0 fp16-matrix

Or run a comma-separated list of scenarios:

vkpeak.exe 0 fp16-matrix,int8-matrix,copy-d2d

You can also explicitly request all scenarios:

vkpeak.exe 0 all

Available scenario names:

• fp32-scalar
• fp32-vec4
• fp16-scalar
• fp16-vec4
• fp16-matrix
• fp64-scalar
• fp64-vec4
• int32-scalar
• int32-vec4
• int16-scalar
• int16-vec4
• int64-scalar
• int64-vec4
• int8-dotprod
• int8-matrix
• bf16-dotprod
• bf16-matrix
• fp8-matrix
• bf8-matrix
• copy-h2h
• copy-h2d
• copy-d2h
• copy-d2d

If you encounter a crash or error, try upgrading your GPU driver:

• Intel: https://downloadcenter.intel.com/product/80939/Graphics-Drivers
• AMD: https://www.amd.com/en/support
• NVIDIA: https://www.nvidia.com/Download/index.aspx

Build from Source

1. Clone this project with all submodules

git clone https://github.com/nihui/vkpeak.git
cd vkpeak
git submodule update --init --recursive

2. Build with CMake

• You can pass -DVulkan_LIBRARY=<path to your macOS/lib/MoltenVK.xcframework/macos-arm64_x86_64/libMoltenVK.a> option to link static MoltenVK library on MacOS, MoltenVK is part of Vulkan SDK from https://vulkan.lunarg.com/

mkdir build
cd build
cmake ..
cmake --build . -j 4

Sample

NVIDIA RTX5060Ti 16GB

device       = NVIDIA GeForce RTX 5060 Ti

fp32-scalar  = 17137.46 GFLOPS
fp32-vec4    = 16910.07 GFLOPS

fp16-scalar  = 12730.03 GFLOPS
fp16-vec4    = 12715.02 GFLOPS
fp16-matrix  = 101485.35 GFLOPS

fp64-scalar  = 398.59 GFLOPS
fp64-vec4    = 394.08 GFLOPS

int32-scalar = 12703.68 GIOPS
int32-vec4   = 12181.98 GIOPS

int16-scalar = 12690.05 GIOPS
int16-vec4   = 12208.29 GIOPS

int64-scalar = 3104.59 GIOPS
int64-vec4   = 2666.86 GIOPS

int8-dotprod = 16101.59 GIOPS
int8-matrix  = 202947.80 GIOPS

bf16-dotprod = 0.00 GFLOPS
bf16-matrix  = 0.00 GFLOPS

fp8-matrix   = 0.00 GFLOPS
bf8-matrix   = 0.00 GFLOPS

copy-h2h     = 18.17 GBPS
copy-h2d     = 17.93 GBPS
copy-d2h     = 18.09 GBPS
copy-d2d     = 190.70 GBPS

AMD RX9060XT 16GB

device       = AMD Radeon Graphics (RADV GFX1200)

fp32-scalar  = 17606.54 GFLOPS
fp32-vec4    = 12155.22 GFLOPS

fp16-scalar  = 16921.16 GFLOPS
fp16-vec4    = 27833.48 GFLOPS
fp16-matrix  = 105337.66 GFLOPS

fp64-scalar  = 442.80 GFLOPS
fp64-vec4    = 437.55 GFLOPS

int32-scalar = 2804.59 GIOPS
int32-vec4   = 2796.74 GIOPS

int16-scalar = 15034.62 GIOPS
int16-vec4   = 26356.38 GIOPS

int64-scalar = 932.14 GIOPS
int64-vec4   = 768.53 GIOPS

int8-dotprod = 53893.32 GIOPS
int8-matrix  = 194476.41 GIOPS

bf16-dotprod = 24427.68 GFLOPS
bf16-matrix  = 105099.82 GFLOPS

fp8-matrix   = 205061.72 GFLOPS
bf8-matrix   = 208234.02 GFLOPS

copy-h2h     = 21.05 GBPS
copy-h2d     = 21.17 GBPS
copy-d2h     = 23.70 GBPS
copy-d2d     = 145.23 GBPS

Other Open-Source Code Used

• https://github.com/Tencent/ncnn for fast neural network inference on ALL PLATFORMS