vkpeak

What vkpeak is

vkpeak is free, MIT-licensed software that uses synthetic Vulkan workloads to measure peak device capabilities. It covers floating-point arithmetic, integer arithmetic, matrix operations, and memory-copy bandwidth. Scenarios include scalar and vector tests for several data types, matrix and dot-product operations, and transfers between host and device. Users can run every scenario or choose one scenario or a comma-separated list. An optional device ID selects the target Vulkan device; the default value, 0, selects the default device. Executables are offered for Intel, AMD, Nvidia, and Apple GPUs, with Linux, macOS, and Windows support. The project can be built from source with CMake after cloning its submodules, and its macOS instructions include linking the static MoltenVK library. The README states that vector-based peak measurements are not real-world workloads. It credits Tencent’s ncnn as other open-source code used. If a run crashes or reports an error, the guidance is to upgrade the GPU driver. The January 2026 Linux package supports glibc 2.17 and later.

Who it suits

vkpeak suits developers and system users who need command-line Vulkan peak-capability measurements. It is useful when synthetic arithmetic, matrix, and transfer scenarios are the intended workload.

Limits

vkpeak reports synthetic peak metrics produced with vector operations, so its results do not describe real-world workloads. The Linux package in the January 2026 release requires glibc 2.17 or later.

Getting started

1. Open the vkpeak repository at https://github.com/nihui/vkpeak
2. Use a provided executable for your Intel, AMD, Nvidia, or Apple GPU
3. Or clone the repository with its submodules
4. Build from source with CMake
5. On macOS, enable linking of the static MoltenVK library if needed
6. Run all scenarios or select one or a comma-separated list
7. Optionally set a device ID; 0 selects the default Vulkan device

Verdict

vkpeak provides configurable Vulkan benchmarking across major desktop operating systems and GPU vendors. Its results describe vector-based peak capability rather than application performance.