My opinion is that it's as reliable and less expansive to use a single widely used GPU (like gaming ones, nVidia 2080) with 4k60 => 4xFullHD 60 splitters like Datapath FX4 or Matrox QuadHead2Go. Distributing the outputs on multiple GPUs will require memory transfers from one GPU to another, and anyway the other GPU is not running since we can not distribute OpenGL processing over multiple GPUs. So better having a single GPU and splitters, spitters will still be useful in 10 years while a GPU is deprecated after 5 years.
I did tests up to 19 screens on a single PC. I had two GPUs but the interresting test is to have 4 datapathes, so 16 fullHD 60 outputs and get them all in sync without frame drops, and as 16 FullHD projectors inside MadMapper:
https://www.youtube.com/watch?v=jwUdfu5LsiU
(Actually my screens are not FullHD but Datapathh is downscaling, so MadMapper handles 19 fullHD outputs through 4 4k outputs to datapath and a TripleHead2Go on a second GPU - don't use those ones, that's a nightmare)
Cheers