I don't know how much you can generalize since the detail level on different maps can be wildly different. But I have done some experimentation myself. I make maps with an extreme amount of detail, and have found that smaller tiles (.img files) perform much better on the GPS. In my case, I try to keep the individual tiles between 1MB and 2MB each. If I use 24k USGS quad boundaries for tiles, it meets this goal for my own maps.
But, as I said, I make maps that have a very high data density, such as this.
http://www.gpsfiledepot.com/maps/view/276/I think the GPS must load all tiles necessary to display your current location on the screen. So you have a trade-off: with large tiles it may only need to access a single tile but it will take longer to load it. With small tiles it will need to load more of them, but they can be accessed more quickly. So let's say you have a map that consists of four 4 MB tiles. If you are right in the middle of that map, the GPS will have to load all 4 tiles, or 16MB of data to cover the screen. But if you make the same map with 16 tiles of 1MB each, the GPS will probably only need to load 4 of those little tiles, or 4MB of data.
Really, it's best to do some experiments on your own and see how it works. Unless you are making a very small map, I don't think that a single .img file will perform very well on the GPS.