The 27 arrangement (right image) looks a lot more natural to me. The symmetry axis is softer, and we have nicer factorisation 3*3*3 (times 3 tables). It even plays nicer with the four channel structure of the amp. A question is how many Pi 3Bs we need to have uninterrupted overwriting of phrases.
Performance test: Pi 3B (not-plus), runs last version of Mellite workspace at around 66 seconds per iteration. With 3 Pis responsible for the entire circle, each Pi has to render 3 buffers, so a rotation cannot be faster than 66*3 = 198 seconds, and thus the minimum phrase length would be 22 seconds if we want to avoid gaps. Ideally, we could come to down to around 15 seconds, so we would need to try to speed up the iteration to no more than 45 seconds :-/
With three Pis, we could have a fixed channel / relay wiring, so that's a lot simpler. If we add a forth Pi, we need a coordinator Pi, essentially a fifth Pi, and then weirdly the memories / buffers would no longer correspond to static angles.
Another idea would be to measure the performance of a Pi 3B+; perhaps we have 4 Pis for each table, the Pi 3Bs responsible for two channels each, the Pi 3B+ for three channels, then we'd use the four amplifier channels, and slightly asymmetric relays.
Another idea would be to try to render two buffers in parallel, making use of the multi-core feature. But this could run into memory problems...
Possible performance improvements: Trying to run headless? Certainly without meters etc. Using a RAM disk for rendering target and/or temp files. Disable VNC etc. I also suspect Akka start/stop taking a bit of time, so ideally there would be the 'reset' feature in FScape, but that is unlikely to materialise in such short time.