Apple has integrated Scalable Matrix Extension (SME) support into its M4 chip generation, allowing matrix operations directly from native code without GPU queuing. This implementation enables significant performance gains in audio processing tasks.
Testing on a Mac mini M4 with 16 GB of RAM showed that processing a bank of 65,536 resonators with SME reduced median calculation time from 1.59 ms to 0.44 ms. At a 48 kHz sample rate and a 64-sample buffer, this allowed the system to meet strict real-time deadlines consistently.
In a separate test, SME outperformed NEON and Accelerate frameworks for resonator bank calculations. For a neural network audio layer with 128 channels, SME proved approximately 12 times faster than Eigen and 1.4 times faster than Accelerate. A full GeneratorNSF vocoder implementation saw a 1.6-fold speed increase.
The developer noted that the SME-based approach maintains high precision, with an SNR of 98.73 dB for resonator banks and up to 119.7 dB for vocoder synthesis. These optimizations allow for more complex timbres or additional processing within the same buffer constraints.
