The Dreams approach certainly has benefits but performance isn't one of them. You say huge worlds but so far we've only seen small scenes (AFAIK). As you say it's noisy and doesn't support the standard PBR workflow. It works for what it's trying to accomplish but I wouldn't recommend it for the majority of games.
That's not quite accurate. Performance of splatting is totally dynamic. You can think of it as a form of rasterization but with built in LOD mechanism.
In Dreams they do this by generating points at the surface with uniform distribution but irregular. At distance they just skip over a subset of points. (Think of instead drawing triangles, draw only the texels of its texture as points, and at distance you go up one mip level so the number reduces to 1/4th)
They have videos where people copy paste parts of the scene to build a city in very short time, and there seems little FPS drop. (I'm puzzled by that myself, because i doubt they use any form of hidden surface removal - just insane brute force compute power?)
Having no support for LOD ist really the main limitation of triangle based rasterization and RT. Computer graphics is mainly about two problems: Visibility and LOD. The rise of GPUs has put the latter out of our attention a bit, but it is still important, especially if we aim for GI in realtime.
This is the major point why i am not so convinced about fixed function hardware, because LOD is a open problem everywhere and solving it always disagrees with FF HW.
PBR works with splatting without any issue. I only meant for Dreams they did not do it because the content is made by the players and they would not like to place environment probes manually, and the devs have very different goals for the artstyle anyways (painterly).
If you read the paper, the programmer has experimented with all kinds of awesome high tech, but it was the artist that has pushed him to do 'boring' splatting, which gave the artistic results finally as intended.
I agree it's not meant to replace triangles for games yet, but at increasing detail levels it would beat triangle rasterization at some point without any doubt. (A graph of triangles is a very complex data structure in comparison to a point hierarchy, but both are just an approximation.)
I expected rasterization HW to become deprecated and finally removed from GPUs. Just compute would remain, and texture filters of course. No limitations. I still think it will happen this way and also RT cores will disappear again... but i see it will take much longer than i hoped for
I might sound unrealistic here, but we made raster HW to put triangles on screen fast. Now this is the smallest problem we have. For GI we need to 'render' the scene from any point, not just for the eye. We have a very different problem now than we have had 20 years ago.