Skip to main content

GPU Acceleration

Configure GPU acceleration for optimal performance.

StemX uses ONNX Runtime for neural network inference, which supports multiple GPU acceleration backends. GPU acceleration makes separation much faster than CPU, especially for longer files. On Apple Silicon, the model runs through MLX faster than real time.

Supported Backends

CUDAFastest

For NVIDIA GPUs (GTX 1060 or newer). Provides the best performance.

Requirements:

  • NVIDIA GPU with CUDA compute capability 6.0+
  • NVIDIA driver version 525 or later
  • Windows 10/11
DirectMLWindows

Works with any DirectX 12 compatible GPU (NVIDIA, AMD, Intel).

Requirements:

  • Windows 10 version 1903+ or Windows 11
  • DirectX 12 compatible GPU
  • Up-to-date GPU drivers
CoreMLmacOS

Automatic on Apple Silicon Macs (M1 and later). Uses the Neural Engine. On Apple Silicon, StemX also uses an MLX backend for the fastest separation.

Requirements:

  • macOS 12 Monterey or later
  • Apple Silicon (M1, M2, M3, etc.) or Intel Mac with Metal
CPUFallback

CPU processing is always available as a fallback. Optimised with AVX2/AVX-512 SIMD instructions where available.

Configuration

Open StemX settings (gear icon) to configure the execution provider:

Auto (Recommended)- Automatically selects the best available backend
GPU Only- Forces GPU processing (fails if unavailable)
CPU Only- Disables GPU acceleration

Performance Comparison

GPU (CUDA / DirectML)Fastest
Apple Silicon (MLX / CoreML)Faster than real time
CPU (fallback)Slower than real time

Actual speed depends on your hardware, the file length, and the buffer size.

GPU Not Detected?

  • Ensure GPU drivers are up to date
  • Check that your GPU meets minimum requirements
  • On Windows, verify DirectX 12 is installed
  • Restart your DAW after changing GPU settings
  • Check the StemX log file for detailed error messages