At GTC20204, Jensen just revealed the new Blackwell platform with an AI superchip with 208B transistors!!! WOW! Imagine what a trillion parameter scale generative AI could achieve?
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Nvidia announces new AI superchip with 208B transistors!!
- Thread starter diplomat33
- Start date
DanCar
Active Member
Yeah, I am sure they are not cheap. But these chips are meant for massive AI training.
DanCar
Active Member
Suspect $40K is just the chip you can't buy.
Update: Toms Hardware confirms the suspicion:
Nvidia's Jensen Huang says Blackwell GPU to cost $30,000 - $40,000, later clarifies that pricing will vary as they won't sell just the chip
Nvidia's GB200 costs $6,000 to make, according to one analyst.
www.tomshardware.com
Wrapping it in something useful cost another $10K?
Complete system $60K?
Some specs, not very interesting for most:
Last edited:
DanCar
Active Member
Theoretical TFLOPS table comparison that includes the just announced B100 and B200
github.com
ml-engineering/compute/accelerator at master · stas00/ml-engineering
Machine Learning Engineering Open Book. Contribute to stas00/ml-engineering development by creating an account on GitHub.
github.com
The only thing that's massive is the bubble AI is in.
Xenoilphobe
Well-Known Member
Don’t forget the nuclear power station you need to power any reasonably sized system built on them.
And I think that really is a thing. But was for cryptocurrency if I recall correctly. Which I seldom do.
Last edited:
It was sarcasm, blended with truth as building these things is my day job. Power is getting out of control now.
Theoretical TFLOPS table comparison that includes the just announced B100 and B200
ml-engineering/compute/accelerator at master · stas00/ml-engineering
Machine Learning Engineering Open Book. Contribute to stas00/ml-engineering development by creating an account on GitHub.
Historically people use FP64 for TFLOPS, so take the numbers with a mine of salt. They’re big, but they’re TF32, not even FP32.
Then there is the matter of keeping it busy all the time, which is programmatically difficult. You also have to use CUDA to get that level of performance in any sensible way.
Then there is the fact that not all applications are suited to be run massively parallel or use tensor cores.
You could keep an A100 mostly busy, people struggle to keep h100 busy unless very specialist. That’s ignoring h200 and never mine the Blackwell.
It’s not about stats in this game, it’s about usability and efficiency for the applications and they are not just AI.
It’s all feeling a bit late 90s to me, with some useful truths and benefits and some crazy hype too.
Similar threads
- Article
