Hi Uli --
This is not an easy answer. Unless there is a reason to deep dive the Tyros 5 (the old era), I'll pass on describing the T5 for now.
The remainder of this answer uses the Montage as reference. The Genos has bigger wave memory capacity and may use different NAND flash integrated circuits than the Montage. Without a service manual, one cannot know with certainty.
ONFI passes the NAND flash address in byte sequential fashion. (There is an option to pass data 16 bits at a time, but the Montage uses 8-bit byte sequential.) The number of bytes transferred is determined by the number of address cycles. The Montage NAND flash is a Cypress S34ML08G101TFI000 which implements 5 address cycles.
With ONFI, there's no guarantee that all address bits of the 40 possible bits are actually used. The S34ML08G101TFI000 implements address bits A0 to A30 for a total of 31 address bits. That's for a single IC (one channel).
The Montage design moves 16-bits at once by operating two S34ML08G101TFI000 ICs in tandem (parallel). So, one question regarding Yamaha's parameters, are the stated capacities 8-bit bytes or 16-bit words? The tone generator operates on two 16-bit words at a time and has two independent channels (HIGH and LOW), moving 32-bits total per cycle.
Thanks to that crazy polyphony spec, I think the two TGs are fed by separate independent wave memories. That makes it a little easier to figure out what's happening on the expansion side (TG#2).
The big assumption here is a low-level hardware design like the Montage and the use of similar ICs for flash.
Yikes. Gotta go to the gym where I will be thinking some more.
-- pj