Product docs
NVIDIA GPU & CUDA Compatibility
Choose CUDA vs Vulkan on Windows NVIDIA GPUs, supported GPU generations, and driver tips for Audio Note Whisper acceleration.
NVIDIA GPU & CUDA Compatibility
Performance
GPU and runtime settings screenshot
What this page is for
On Windows with NVIDIA GPUs, Whisper can use Vulkan or a CUDA engine. Newer cards often work with both; GTX 10-series and older cards are more likely to crash or fail to initialize on Vulkan—the app will steer you toward CUDA.
Read this before relying on GPU acceleration or debugging engine errors.
Which NVIDIA GPUs (user-facing)
| Generation | Examples | Guidance |
|---|---|---|
| GTX 10-series and newer | 1050 Ti, 1060, 1070, 1080, … | Prefer CUDA; Vulkan may be unstable |
| GTX 16 / RTX 20–50 series | 1660, 2060–5090, … | Often CUDA or Vulkan—benchmark on your machine |
| Older than 10-series | 900-series and below | GPU acceleration may be unreliable—use CPU or newer hardware |
Non-NVIDIA GPUs: Vulkan or CPU—see GPU Transcription.
CUDA vs Vulkan
| Situation | Pick |
|---|---|
| App says Vulkan crashed or GPU incompatible | Switch to CUDA (Settings → Transcription → GPU engine) |
| First CUDA use | Download engine + runtime packages (hundreds of MB)—follow in-app steps |
| Clean setup on a new card | Try Vulkan first; move to CUDA if unstable |
| Mostly realtime models | GPU engines are usually not required for realtime models |
If in-app messaging conflicts with this page, trust the app.
Drivers and environment
- Keep NVIDIA drivers reasonably current.
- After a failure: update driver → reboot → re-download GPU packages in Settings → compare CPU vs GPU on a short clip.
- If GPU still fails, stay on CPU and note GPU model + engine when contacting support.
FAQ
CUDA download is large—can I skip it?
Yes; use Vulkan or CPU until ready. Legacy NVIDIA users should finish CUDA setup before relying on GPU.
Standard vs Pro for GPU?
GPU toggle is available on Standard and Pro (Free is CPU-only); plan differences are mostly models, realtime tuning, and recording.
Does this apply to advanced ASR or realtime models?
They use separate engine and model downloads—not the Whisper CUDA/Vulkan packages.
