> ## Documentation Index
> Fetch the complete documentation index at: https://docs.echophrase.com/llms.txt
> Use this file to discover all available pages before exploring further.

# GPU Troubleshooting

> Fix DirectML and GPU acceleration issues

# GPU Troubleshooting

Solutions for GPU acceleration problems.

## GPU Not Detected

Echophrase auto-detects the fastest available GPU at runtime. If
**Settings → GPU Acceleration** shows "Using CPU" on a machine with a
DirectX 12 GPU, work through the checks below.

### Update your GPU drivers

Most GPU detection problems come down to outdated drivers. Install the
latest from your vendor:

* **NVIDIA**: [GeForce / Studio drivers](https://www.nvidia.com/Download/index.aspx)
* **AMD**: [Adrenalin](https://www.amd.com/en/support)
* **Intel**: [Arc / Iris drivers](https://www.intel.com/content/www/us/en/download-center/home.html)

After updating, restart your machine and relaunch Echophrase.

### Confirm DirectX 12 support

DirectML requires DirectX 12. Press **Win + R**, type `dxdiag`, and check
the **Display** tab — under "Drivers" the **Feature Levels** line should
include `12_0` or higher. Any GPU released since \~2015 should qualify.

### Already on CUDA?

CUDA support was removed in favor of DirectML — Echophrase no longer
ships CUDA at all. If you arrived here from older docs about installing
CUDA Toolkit or cuDNN: you can uninstall those if they were installed
purely for Echophrase. Update to the latest Echophrase build and the
GPU should be picked up automatically via DirectML.

## GPU Detected But Not Used

Open **Settings → GPU Acceleration** and confirm:

1. **Auto (Recommended)** is selected — this lets Echophrase pick the
   fastest device at session start.
2. The inline status at the top shows your GPU (for example:
   "Active: Using DirectML (NVIDIA GeForce RTX 3080)").

If the status still shows "Using CPU", restart the app after a driver
update. The execution provider is selected once at session start.

## Out of Memory Errors

If you see GPU out-of-memory errors:

1. Use a smaller model — switch from **Parakeet-TDT 0.6B** to **Parakeet
   Lite** (see [Requirements](/requirements) for the full matrix)
2. Close other GPU-intensive applications
3. Restart Echophrase to free GPU memory

## Performance Issues

If GPU transcription is slow:

1. Confirm the active provider is DirectML (or NPU on Copilot+ PCs), not CPU
2. Make sure your GPU isn't thermal throttling under load
3. Try Parakeet Lite for faster results on lower-end hardware