By my calculations we need to have 90 - 120hz at 16000x9000 per eye in order to have full retina level PPD of about 120ppd vs the 8KX’s 25PPD~.
8KX requires 1,492,992,000 pixels per second at 90hz refresh and 3840x2160 resolution X2 eyes.
A 32K would require 31,850,496,000 pixels per second at 120hz refresh and 15300x8640 resolution per eye X2 eyes, which is about 21x as much processing power as the 8KX.
Assuming the 4090 is 2.5x more powerful than the 3090 and the 3090 can manage 90hz on the 8KX, and performance doubles every time NVIDIA doubles the GPU tiles from 1 to 2 to eventually 4, then NVIDIA cards could manage the 32K at 60fps at 4 tiles.
Assuming performance doubles again with a 512 bit bus and once more with advancements from GDDR7/8 and using Infinity fabric /3d stacking of processor cache memory we may actually have the performance required to push 30 billion pixels per second by 2035-2040.
This isn’t even factoring in performance improvements from node shrinks from 7mm down to 0.7mm by the end of TSMC’S roadmap.
It’s not factoring in software and firmware updates either which will improve performance as rendering efficiency methods further improve. One issue is I dont think a cable will exist anytime soon that can deliver 20x the bandwidth of display port 1.4. (there needs to be a cable capable of 680Gbps data transfer)
Performance still on the table
512 bit bus and GDDR8 - 2x
Double GPU tiles - 2x
Quadruple GPU tiles - 2x
Node shrinks from 7mm to 0.7mm - 2x
3D stacking/infinity fabric technology - 2x
Rendering method improvements/driver/firmware - 2x
I think we only need to see a doubling of performance 3 times from the 4090 to run the 32K if 4090 is 2x as fast as the 3090, and even though we have up to 6 times doubling of performance left on the table. Should be enough performance on the table to run Unreal engine 5 games in VR at this rate. Hopefully around the 7090/8090 period which is going to be around late 2028.
In terms of standalone, the 7000 series ryzen APUs perform as good as a 3060M which is equivalent to a 1080s performance, but in an integrated chip
Kevin mentioned that foveated render can reduce render requirements by a lot, say would we have the performance to run 32k between both eyes in around 5-10 years when we have 4-8 tile die GDDR8 powered GPUs with 512 bit buses made on 0.7nm TSMC 3d processes? I think it could be possible to see this kind of performance, what do you guys think?