Given the push towards job based systems, and the fact that GPU processing is already massively parallel, it should be theoretically possible to scale to many GPUs, well over 4. The limitation will likely be related to how much overhead (both actual processing time and developer man hours) it takes to split the tasks to such a small level of granularity and distribute them. My guess is only the big engines are going to have explicit multi adapter, less of those will have iGPU supporting multi adapter, and even less than those (perhaps just one or two) will have multi adapter support exceeding 4 GPUs.