maybe
Albert Penello would be able to at least shed some knowledge on this hypothetical. which would be less expensive?
If I understand the question posed, the idea is if it's cheaper to run an existing chip at a higher spec (therefore having less usable parts) to hit a performance target (say 10 TFLOPS) vs. building a new chip from scratch that has overhead to disable CU's to also get to 10 TFLOPS?
Is that the gist?
This is a super interesting question I'll have to think about it. There are a lot of variables at play, not the least of which is -
when is this decision being made?
In general, the answer would doubtless be to build the right chip. Getting good yields and having overhead is going to be the most cost effective in the long run. That said, the sunk costs for a whole new chip development are, like, A LOT. So you don't want to do that twice.
But really it's sort of unanswerable without a lot more information. I think it really comes down to how much the yields are impacted and how the longevity of the good chips are affected. I've said before that +/- 10% is a good rule of thumb where you could consider the performance increases vs. the yield implications to be tolerable.
At the risk of beating a dead horse (since I've said this often), I also need to point out the case and cooling design limitations. You can can crank the silicon all the way up until you only have 1 good chip off the wafer, but if you can't cool it once it's in the console, it's no good. So that's a limiting factor in this as well.
(QUICK EDIT: Remember that the cost of throwing away bad chips goes up at a faster rate then the number of bad chips. So if you were expecting 90% yields, removing 20% more chips is ~30% more expensive per chip. Removing 30% more chips makes each chip 50% more expensive, and so on)
