New in v2.6.1.2 (experimental)
compute_backend.py allows hardware acceleration for neural calculations by changing one line:
in config.ini:
[Compute]
backend = numpy
Options:
-
numpy- default, no extra dependencies -
onnx- enables hardware AI accelerator support via ONNX Runtime
auto-selects DirectML, OpenVINO, QNN, or falls back to numpy if no runtime is present.
requires runtime to be installed (refer to the following list:)
Recommended ONNX runtime packages by platform:
-
Windows | NVIDIA + AMD + Intel GPU + NPU (DirectML) |
pip install onnxruntime-directml -
Windows | NVIDIA only (maximum CUDA performance) |
pip install onnxruntime-gpu -
Windows | Qualcomm 8CX / SQX / Snapdragon (NPU) |
pip install onnxruntime-qnn -
macOS | Apple Silicon and Intel Macs |
pip install onnxruntime -
Linux | NVIDIA GPUs |
pip install onnxruntime-gpu -
Linux | AMD GPUs | Use the ROCm/MIGraphX build (see ONNX Runtime docs)
If no runtime is installed, the default `numpy` will be used (neural calculations performed on CPU)
ONNX support is NEW and EXPERIMENTAL - Please report issues!
The Brain Tool About Tab displays which backend is being used: (version 2.6.2.0+)
