Denial:

Nvidia already has inference chips utilizing chiplets in testing. They also published a number of papers that talk about chiplets being the only way to get scaling going forward. Bill Dally's (Nvidia) recent interview he says they consider chiplets on GPUs "de-risked" and ready for actual implementation. There has already been rumors that Hopper (next gen after Ampere) is going to utilize chiplets. That being said it's completely different then multi-GPU and it doesn't seem like it's going to be good for gaming. It's highly likely we'll only see it in HPC GPUs for sometime across all three GPU designers.

Even using chiplets is not the answer to ongoing improvements, because the amount of chiplets that can be used in a package will not be able to double every 2 yrs to replace Moores Law...I see it as a short/medium term stop gap to help fill in deficiencies in Moores Law whilst die shrinks and transistors per mm progress has been slowing down...the end is nigh (silicon)!

Andrew LB:

I had thought i'd die of old age before Intel brings a proper GPU to market. heh... who knows though, still might.

This. Let me know when Intel has a gpu that directly competes with Nvidia and AMD for desktop gaming.

To quite intel: "you can't just glue CPU dies together, the performance is terrible" *proceeds to do exactly that, as they did 15 years ago*