It's bigger! Bigger and more power efficient. Ok the doubling, again, of 8bit performance is cool. Something neat there I suppose; though I feel it's odd to put this massively expensive card out there for... inference work. Who'd buy it for that? Well it'll help on their next gen of inference specific hardware. But otherwise, it's just bigger-er. And that's fine and dandy for their "we can charge a hell of a lot per card" AI accelerators.
But other than limiting itself to a standard power input, again putting a total damper on the "850 watts 4090" stuff, it doesn't feel like this announcement gives us a whole lot to go in for the consumer end. I mean surely you can half the size of this massive card, probably about what a 4090 would be, limit it to 600 watts and still have headroom to clock it up to 2.something ghz. As for bandwidth. HBM? Probably not. No word on any changes to the cache structure either, so no hints of a big LLC either way; and the big AI card might not even include such a feature.
Guess, other than concluding the powerdraw could easily stick to 600watts or less, we'll have to wait 5 months or whatever until the Hopper consumer launch to find out.