yes they can and they will
but they will only do so when monolithic designs fail to provide performance gains which it's still doing.
when one generation to the next of graphics card drops under 10% performance boost then we'll see them push up multi core/chip designs
but as of 2019 nvidia and amd has no issues turning more transistors into more performance using existing architecture designs and without turning the gpu into a nuclear plant which is what was happening with cpus that's why they moved to dual core
The problem is with how a game works you have very discrete frames of data, very small timeslices to work with and you can't just go X chip has the data I need *yoink* even with a fast interface between them as you need to know what the other GPU is doing, whether it has finished processing that data, when you can actually fetch it without causing a slowdown on that GPU as you can't just start reading from memory while the other GPU is active and the data will necessarily remain intact while you do so without the other GPU being part of the process - which inhibits its ability to do its job, etc.
Some big improvements are needed to either dynamically pool hardware resources on the fly or better farm out areas of a scene so that you don't have one GPU package sitting there almost idle rendering the sky while the other is working flat out rendering a more complex part of the scene, etc. but that still doesn't address the situations where the effect one GPU might be processing could depend on data from a previous frame that sits on another GPU or data that the other GPU is currently working on, etc. stalling processing until it is available.
There simply isn't a way to simply make something like Ryzen but for GPUs work with game rendering especially not without a complete ground up reworking of how GPUs work.
) but most software can’t?
