GPU acceleration backends

Minutes transcribes with whisper.cpp, which can run on the CPU (default) or on a GPU through one of several backends. GPU backends are opt-in Cargo features.

Feature	Backend	Platforms	Notes
(none)	CPU	all	Default. Works everywhere, no toolchain needed.
metal	Apple Metal	macOS (Apple Silicon)	Fast on M-series; needs full Xcode for the shader compiler.
cuda	NVIDIA CUDA	Linux, Windows	Needs the CUDA Toolkit + a supported GPU/driver. See the Pascal + Windows caveat below.
vulkan	Vulkan	Linux, Windows	GPU acceleration with no CUDA Toolkit. The most portable NVIDIA path on Windows. Needs the Vulkan SDK (LunarG).
hipblas	AMD ROCm/HIP	Linux	For AMD GPUs.

Install by adding a GPU feature to the defaults, for example:

cargo install minutes-cli --features vulkan

Do not pass --no-default-features with only a GPU feature. whisper is a default feature; dropping the defaults builds a binary with no transcription ("Transcription placeholder — whisper feature not enabled"). If you must use --no-default-features, name whisper explicitly: cargo install minutes-cli --no-default-features --features whisper,vulkan.

Choosing a backend

• Apple Silicon: use metal.
• NVIDIA on Linux: cuda if your Toolkit + GPU generation are supported, otherwise vulkan.
• NVIDIA on Windows: prefer vulkan. It avoids the CUDA/nvcc/MSVC toolchain entirely, which is where most Windows build failures live (see below).
• AMD on Linux: hipblas.
• Unsure, or the GPU build fights you: the default CPU build always works. GPU is a speed optimization, not a requirement.

Known limitation: CUDA on Pascal (GTX 10-series) + Windows

Building --features cuda on Windows for a Pascal GPU (compute capability 6.1 / sm_61, e.g. GTX 1050/1060/1070/1080) is not viable with current toolchains, and there is no minutes-side fix because the failure is inside NVIDIA's proprietary compiler, before any Minutes or whisper.cpp source compiles:

• CUDA 13.x dropped Pascal support outright (nvcc fatal: Unsupported gpu architecture 'compute_61').
• CUDA 12.6 (the last Toolkit with full Pascal support) crashes during CMake's compiler-identification step: nvcc error: 'cudafe++' died with status 0xC0000005 (ACCESS_VIOLATION). cudafe++ is NVIDIA's closed-source CUDA frontend; the crash is on the exact invocation CMake constructs, and nothing in the Minutes build controls it.
• CUDA 11.8 fails to compile CMake's probe against the modern MSVC STL (STL1002 and residual parser errors from nvcc 11.8's older frontend).

This is a dead toolchain combination, not a bug we can patch — cudafe++ is not open source, and the crash precedes whisper.cpp entirely, so neither a CMake nor a whisper.cpp change would fix it.

Try Vulkan instead — but see the shader-gen caveat below:

# Install the Vulkan SDK (LunarG) first, then:
cargo install minutes-cli --features vulkan --force

Vulkan avoids the CUDA Toolkit and nvcc entirely, sidestepping every failure above. Keep the default features (do not pass --no-default-features with only vulkan, or you get a placeholder binary with no transcription — see the note in the install section above).

Vulkan also fails on some Windows configs: vulkan-shaders-gen hangs

On at least one Windows machine (a dual-GPU laptop: Pascal GTX 1050 + Intel UHD 630, Vulkan SDK 1.4.350), the vulkan build hangs indefinitely in whisper.cpp's ggml-vulkan step: vulkan-shaders-gen.exe sits at 0% CPU forever (reproducible serial and parallel, and by running the extracted command manually outside the build). This is an upstream ggml/whisper.cpp build-tool hang, not a Minutes issue, and it is under investigation upstream (issue #531). A 0%-CPU block (rather than a spin) points at the tool waiting on a subprocess/device call, and dual-GPU device enumeration is a suspected factor.

If both CUDA and Vulkan fail on your machine, use the CPU build — it always works, needs no toolkit, and on a low-end GPU like a GTX 1050 the GPU speedup is modest anyway:

cargo install minutes-cli --force

Reference: issue #531.