Choose precision for the workload | What was compared? | Selected full-benchmark results | Small workloads can be faster without CUDA | Why can the native pipeline be faster? | Installed footprint | When cudaverse is a good fit | Reproduce and interpret the evidence

Platform support | 1. Install and verify the NVIDIA driver | 2. Install CUDA 12.x | Windows | Linux | macOS | 3. Install cudaverse | 4. Check that cudaverse can use CUDA | 5. Run a strict CUDA smoke test | Common problems | nvidia-smi fails | cublas_loaded or cusolver_loaded is false | CUDA is found, but the test calculation fails | CUDA runs out of memory | A graph or embedding result contains a non-CUDA stage

CUDA task summary | Dense CUDA example | Sparse CUDA example | Graph and embedding workflows | The lightweight backend

Before you start | PCA and exact nearest neighbours on CUDA | GPU-resident matrix operations | Sparse normalization, PCA, and kNN | Other CUDA tasks | Where to go next

Require CUDA at the start | Upload once and reuse the tensor | Keep PCA followed by kNN on CUDA | Keep sparse preprocessing on CUDA | Read provenance | Control memory and transfer costs | Confirm the backend used