CUDA acceleration

This guide describes the current build.

Samsara runs on the CPU. An optional NVIDIA runtime pack exists, but it is not built for the current builds and its compatibility is not proven.

Compatibility evidence is limited. Release builds are CPU-only; the CUDA pack is not built by the release process. The only pack is Samsara-CUDA-Pack-v0.20.0.zip (ten DLLs), hash-verified against the CUDA environment used to test the v0.22 application archive. It does not establish compatibility with the current build.

Use CPU transcription unless you can independently validate the pack in your own environment. A compatible NVIDIA GPU is optional; CPU transcription remains fully supported.

What the older tested route required

  1. Close Samsara.
  2. Extract all ten DLLs from the CUDA pack into Samsara\_internal\ctranslate2\.
  3. Open Settings → Dictation → Advanced → Compute device and select CUDA (NVIDIA GPU). The CUDA entry appears only when all ten files are found. In the older build the row is under Settings → Advanced.
  4. Restart Samsara and check the main log for Device: cuda, Compute: float16.

An incomplete runtime set falls back to CPU, and Settings says which DLLs are missing. In the current build, automatic device selection also requires a working CUDA device count, otherwise it uses the CPU. Do not install a random PyTorch build only to obtain DLLs; that is not a compatibility guarantee.

Known quirk. The Compute device list has no “auto” entry. If your configuration says auto, the row shows CPU, and applying Settings may write cpu. Check the row after you apply.