Compute acceleration

This feature is in development and is not part of the current release. It is planned for MusicStation 0.7.0.

Compute acceleration is in development. The engine schedules its own built-in devices across the three compute units of an Apple Silicon Mac: convolution reverb with long impulse responses, physically modelled instruments and large spectral effects run on the GPU through Metal, while stem separation, audio-to-MIDI transcription, tempo, beat, key and chord detection, denoise and neural amp models run on the Neural Engine through Core ML.

The point of the split is headroom: at the same CPU load, more instances of the built-in devices fit into a session, and the analysis jobs that feed the MIDI editor and the browser are finished sooner instead of occupying a core. Everything runs locally on the Mac - the models ship inside the application, nothing is uploaded, and nothing leaves the machine.

Safety comes first: every device keeps a complete CPU path, whether a backend is available or not. When a GPU result misses its deadline, the CPU path computes that block without a dropout, and a device that misses its deadline repeatedly moves back to the CPU on its own. Each device header shows where it runs - GPU, NE or CPU - so the state is visible instead of guessed.