I know it's little bit CPU intensive, but my main aim was to implement a super-accurate DX7 clone so that I can use the original DX7 presets. So my priority went into that. In following weeks, I will try to reduce the CPU consumption, but I don't think it will change the results drastically. Maybe 10 percent or something. Initially, I planned to combine it with Bela for the Eurorack, but I found out this CPU problem too, so I didn't continue on the project.
One solution would be to remove all functions and calculate them before compiling for each preset and each note and each velocity. This would create a big 127x127x127x127 data (if you want to have all presets) but this might reduce the CPU. But it is almost impossible to reduce the components inside the Synthdef's. They are all used in all cases and most probably a significant portion of CPU usage comes from there.