To answer myself, I have seen one very real difference recently.

For a while I was doing a read/modify/write sequence on VRAM (SCB1) values.

Running on MAME, I had no issue setting the VRAM address and then reading data immediately.

On real hardware, this didn't work at all. In fact I had to insert quite large delays (ballast code) between setting the address and reading the data in order to get them to work properly. To be fair, this is documented (in terms of VCLKS) in Charles McDonald's MVS Hardware Notes document (circa 2007). By my calculations, that's a lot of time required before you can read the data!

Obviously this was a pretty silly way (for me) to do it, and I re-worked my logic to eliminate the VRAM reads altogether.

Just wondering whether anyone else had run into this issue before? Or did you all know well enough not to even attempt VRAM reads?