I'm really curious as to why AMD went the route of parking the 2nd CCD instead of doing what people have been doing for a while with multi-CCD chips and pinning the game process to one CCD using a similar method to process lasso. I'd like to think AMD have some special insight but given how their software dev seems to usually happen it's more likely it was just easier this way for them xD
What causes the latency issue in the performance bottleneck is not running games on multiple CCD's, its moving workload between CCD's, Windows loves to do that, i cannot just recognise that work is being done on a core and just leave it be, it loves to move it, and then move it again... and again.......
With Zen 3 AMD added something that stops windows doing that, Windows wants to move something from core 2 to core 12 and whatever AMD did stops it.
This is different again because now a game needs to run in just one CCD, the one with the 3D Cache, they added something to the drivers to make that happen, tho its not perfect, there are instances where it will run on the CCD without the 3D cache, tho its rare, the only instance i can find of it is the screenshot below, the only way around it is to turn the CCD without the 3D cache off, you can see that in the slide (7800X3D simulated vs 7950X3D) and the difference its made.