- Vulkan supported videocards
- Windows 10/11 x64
By default, whisper.cpp builds typically run on CPU or CUDA.
If you have an AMD GPU, you can accelerate inference using Vulkan, but this requires compiling whisper.cpp from source with Vulkan support enabled.
This guide explains how to build whisper.cpp with Vulkan support on Windows.
You will need the following software installed:
- Vulkan SDK
- CMake
- Git
- Visual Studio Build Tools (usually already present on most developer systems)
Download the Vulkan SDK from: https://vulkan.lunarg.com/sdk/home
Choose: Windows x64 Installer Example: vulkansdk-windows-X64-1.x.xxx.x.exe
During installation:
- Select Install Everything
- Allow the installer to set environment variables
After installation open a Windows Command Prompt and run:
echo %VULKAN_SDK%
You should see something similar to:
C:\VulkanSDK\1.3.xxx.x
If this path appears, Vulkan is correctly installed.
Open a Windows Command Prompt and run:
winget install Kitware.CMake
Example output:
Found CMake [Kitware.CMake] Version 4.2.3
This application is licensed to you by its owner.
Microsoft is not responsible for, nor does it grant any licenses to third-party packages.
Downloading https://github.com/Kitware/CMake/releases/download/v4.2.3/cmake-4.2.3-windows-x86_64.msi
██████████████████████████████ 34.7 MB / 34.7 MB
Successfully verified installer hash
Starting package install...
Successfully installed
Open a Windows Command Prompt and run:
git clone https://github.com/ggml-org/whisper.cpp.git
Example output:
Cloning into 'whisper.cpp'...
remote: Enumerating objects: 31665, done.
remote: Counting objects: 100% (149/149), done.
remote: Compressing objects: 100% (75/75), done.
remote: Total 31665 (delta 79), reused 74 (delta 74), pack-reused 31516
Receiving objects: 100% (31665/31665), 33.63 MiB | 23.60 MiB/s, done.
Resolving deltas: 100% (23050/23050), done.
Navigate to the source folder:
cd whisper.cpp
Optional: remove previous build directory for a clean build:
rmdir /s /q build
Enable Vulkan using:
cmake -B build -DGGML_VULKAN=1
Some CPUs crash when AVX512 is enabled.
| CPU | AVX512 |
|---|---|
| Intel 12th/13th gen | ❌ |
| AMD Ryzen | ❌ |
| Xeon / Threadripper |
For maximum compatibility disable AVX512:
cmake -B build -DGGML_VULKAN=1 -DGGML_AVX512=OFF
Even with AVX512 disabled ggml still uses optimized CPU instructions:
- AVX2
- FMA
- OpenMP
You will see these enabled in the build log, so CPU performance remains good.
Once configuration is complete build the project:
cmake --build build --config Release
After the build finishes the binaries will be located in:
build\bin\Release\
Example files produced by the build:
whisper-vad-speech-segments.exe
whisper-server.exe
whisper-quantize.exe
whisper-cli.exe
whisper-bench.exe
test-vad-full.exe
test-vad.exe
main.exe
bench.exe
whisper.dll
ggml.dll
ggml-vulkan.dll
ggml-cpu.dll
ggml-base.dll
Older versions of whisper.cpp included:
examples/stream/
This example used PortAudio or SDL for microphone capture.
It is often not built by default anymore because:
- Audio capture APIs differ per operating system
- Many users prefer using FFmpeg pipelines for streaming audio input
Vulkan acceleration can significantly improve performance on AMD GPUs compared to CPU-only inference.
Performance depends on:
- GPU model
- Vulkan driver quality
- Whisper model size (tiny → large)
Benchmark the system:
whisper-bench.exe
Run transcription:
whisper-cli.exe -m models/ggml-large-v3.bin -f audio.wav
Official repository: https://github.com/ggml-org/whisper.cpp
Other Vulkan builds: https://github.com/jerryshell/whisper.cpp-windows-vulkan-bin