When MP/M came out, I think it's a fair estimate that the most common "full up" machine had 64K of RAM; probably 4116 type DRAMs. MP/M would not have done well. There was no standard for bankswitching; everyone did their own thing. For example, on the Durango F85, we ran the top 6 bits of the CPU address through a Fairchild 64x9 bipolar SRAM--translation speed was a big issue as we were using a 8202 DRAM controller.
That gave you up to 256K of memory space (we didn't use the 9th bit), mappable in 1K blocks. It was elegant and we used it to good advantage in our own software. It wasn't what MP/M was designed to use, although it was possible to use it--but then, we didn't design our system to run MP/M (or CP/M). We had our own multitasking OS.
In short, this wasn't the era of standardized conventions, as you see in the PC or Mac world. There were some wild differences between vendors.
For multitasking, there was also the one-CPU-per-user mode, as one might find on a Molecular system. The "master" CPU handled I/O, but otherwise, each user had his own 64K address space. In a bus-oriented system, such as S100 or Multibus, this makes a lot of sense. Again, not something that MP/M was designed to.