I can't wait to see the GPU equivalent of Backblaze's hard disk reliability reports. And, considering Intel is building a 1000W CPU with direct liquid cooling, I would love to see one for CPUs as well.
I wonder if multi-phase cooling, fridge-stye, would be an option - pushing sub-zero fluids to the heat exchangers on top of the chips to remove more heat than just water would.
As a fun side note, IBM mainframe CPUs run at 5.2 GHz continuously for years without significant expected failures. The latest one uses liquid cooling with glicol.
The really interesting thing is to see the difference between Nvidia chips and custom chips from Google, Amazon and others (but we never will).
I bet that Google and Amazon custom chips last that 1-2 years and Nvidia chips lasts 5 years on average.
> I bet that Google and Amazon custom chips last that 1-2 years and Nvidia chips lasts 5 years on average.
Makes sense. They don't need to last much more than one product iteration. By the time they fail, there's already a new platform taking its place.