GPU Acceleration

    Speed up transcription using your device's graphics processor.

    Overview

    GPU acceleration can make transcription significantly faster than CPU-only — larger models benefit most. On capable hardware, a 30-minute meeting can be transcribed in seconds rather than minutes.

    Clearminutes automatically detects and enables your GPU. No manual configuration is needed.

    Check If GPU Is Enabled

    GPU acceleration is enabled automatically when available. You can confirm it's active:

    1. Open the Developer Console (Cmd+Shift+I on macOS or Ctrl+Shift+I on Windows)
    2. Go to the Console tab
    3. Start a transcription
    4. Look for one of:
      • [WHISPER] Metal GPU enabled (macOS Apple Silicon)
      • [WHISPER] CUDA GPU enabled (Windows NVIDIA)
      • [WHISPER] Vulkan GPU enabled (Windows AMD/Intel)
      • If you see none of these, Clearminutes is using CPU

    GPU Acceleration by Platform

    macOS — Metal (Automatic)

    Metal acceleration is enabled automatically on Apple Silicon Macs (M1, M2, M3, etc.). Intel Macs fall back to CPU. No configuration needed.

    Windows — CUDA (NVIDIA)

    CUDA acceleration is enabled automatically when an NVIDIA GPU with current drivers is detected. Clearminutes bundles the CUDA runtime — no separate CUDA Toolkit installation is required.

    If CUDA isn't detected:

    • Update your NVIDIA drivers to the latest version
    • Restart Clearminutes
    • Restart your computer

    Windows — Vulkan (AMD/Intel)

    Vulkan acceleration is enabled automatically for AMD Radeon and Intel Arc GPUs with current drivers.

    If Vulkan isn't detected:

    • Update your GPU drivers to the latest version
    • Restart Clearminutes
    • Restart your computer

    GPU Memory Requirements

    Different Whisper models need different amounts of GPU VRAM. Larger models are significantly faster with GPU but may not fit on older hardware:

    • tiny / base: Work on any GPU
    • small / medium: Most modern GPUs
    • large-v3: Requires a capable GPU (Apple Silicon M1+, or dedicated GPU with 4+ GB VRAM)

    If a model doesn't fit in GPU memory, Clearminutes automatically falls back to CPU.

    CPU Fallback

    Clearminutes automatically falls back to CPU if the GPU is unavailable, runs out of memory, or has an outdated driver. CPU transcription works but is significantly slower.

    Troubleshooting GPU Issues

    GPU shows as "not available"

    macOS:

    • Ensure you're on macOS 10.15+ (check System Preferences → About)
    • Metal is automatic; no action needed
    • If still not working, try restarting the app

    Windows CUDA:

    • Verify nvcc --version works in PowerShell
    • Restart your computer (required after CUDA install)
    • Check NVIDIA drivers are latest version
    • Reinstall CUDA Toolkit if needed

    Windows Vulkan:

    • Update GPU drivers to latest
    • Restart Clearminutes
    • Restart your computer

    GPU is enabled but transcription is slow

    Possible causes:

    • GPU memory is full (close other apps)
    • GPU driver is outdated (update to latest)
    • Large model overloading your GPU (try smaller model)
    • GPU sharing with screen rendering (common on older laptops)

    Try:

    1. Close other applications
    2. Use a smaller Whisper model
    3. Update GPU drivers
    4. Switch to CPU mode and compare speed

    GPU transcription fails partway through

    Usually caused by:

    • GPU ran out of memory (too many large apps open)
    • GPU overheating (clean dust from vents, ensure airflow)
    • Outdated GPU driver (update to latest)

    Fix:

    1. Close large apps (Chrome, video editing software)
    2. Restart Clearminutes
    3. Use a smaller Whisper model
    4. Update GPU drivers

    CUDA/Vulkan won't install

    Windows:

    • Ensure you have admin rights
    • Restart your computer (some installs require this)
    • Try uninstalling and reinstalling
    • Check you have 10+ GB free space (CUDA Toolkit is large)

    Performance Tips

    Maximize Transcription Speed

    1. Enable GPU (Metal, CUDA, or Vulkan)
    2. Enable VAD (filters silence, reduces processing by ~70%)
    3. Use smaller model if possible (tiny → base → small)
    4. Close other applications (frees up GPU memory)
    5. Close browser tabs (reduces system load)

    Balanced Setup

    • Model: base or small (good accuracy, reasonable speed)
    • GPU: Enabled (use system GPU)
    • VAD: Enabled (medium-high sensitivity)

    Highest Accuracy

    • Model: large-v3 (best accuracy)
    • GPU: Enabled (necessary for speed)
    • VAD: Low sensitivity (keeps more audio)
    • Microphone: High-quality USB microphone