the latency of MG is mostly ruled by the wavelength of the notes, and less by the buffersize used. Going lower than 64 samples is mostly not advisable. MG has a latency of 20ms and up, independent from the buffersize.
Splitting one specific process over more cores will mostly make the latency larger, because the process has to have its jobs split, and a ptich to mdii convertor can not be done in parallel processes.
