140 billion transistors for Nvidia Hopper GH100 GPU?
Click here to post a comment for 140 billion transistors for Nvidia Hopper GH100 GPU? on our message forum
Denial
In datacenter it's all about performance density, so this doesn't surprise me. Really interested to see a deep dive into the architecture though - I think we typically get those around March historically so hopefully that's the same this year. Seems like Nvidia's datacenter parts are going to be much different than gaming ones from here on out.
JamesSneed
Maybe if its multiple dies even if they are directly connected on silicon like Cerebras. Not like Nvidia makes its own chips so we would all know if TSMC or Samsung figured out how to increase the reticle limit(they didn't).
rl66
Do you mean the minning one?
More seriously, the compute/IA line despite being based on gaming GPU have already taken this path, so it would be logical to go even more this way.
cucaulay malkin
same as GV100 to GA100,that was 2.56x too
same two years between them too.
another 60% performance increase for +100W power ? seems likely.
I'm not complaining,will settle for a "4070" ~3080/Ti equivalent w. 250W TDP and I'll be happy if it's gonna be available at 650eur.I could really use that +50% performance for rt.
Silva
3090 is 628 mm²
Who the f will afford a 1000mm² GPU? (Don't answer miners, I mean real people)
Just stop making larger and larger chips only 1% of the world can buy!
Krizby
Why not? as long as there are "affordable" GPU being made (with the help of Intel), why should you care about Nvidia making expensive GPU for the 1%?
You know by selling higher margin products, company can afford bigger production capacity and in turn able to produce cheaper products (economy of scale), prime example would be Tesla.
cucaulay malkin
there will be no 1000mm2 gpu
and hopper is for HPC actually,so sorry if I confused people with my earlier post.
that said,2080Ti was 750mm2. And still 999msrp. Same as 520mm2 6900xt.I think there's more to end consumer cost than the die size alone.
The Goose
I feel for the poor guy that counted them
MonstroMart
Legends say he's still counting. The 140 billions figure is just his best guesstimate based on his current progress.
NCC1701D
A dollar per transistor?
asturur
This is a data center computing product, later we will hear from them about graphic unit.
Also if this was 1 9k$ gpu, that would be of course their flagship marketing product, if you want to be the top dog and bully the competition, you have to demonstrate also technological superiority even if is not market viable
alanm
What will make it extra expensive is that the larger the die, the poorer the yields. Flaws more likely to occur in larger dies.
