Nvidia reveals Blackwell B200 GPU, the 'most powerful chip' for AI | Hacker News Reader

Nvidia reveals Blackwell B200 GPU, the 'most powerful chip' for AI

theverge.com

24 pointsmfiguiere17 comments

The article makes no mention at all of how much VRAM they might have. Other articles state plans for up to 192GB of HMB3e, and a power draw of 1000 watts.

I know it’s only FP4, but 1.4 exaflops in one rack is still crazy.

With scaling and some tricks, FP4 inference can be very close to 16 bit.

And most software out there is still missing on FP8 lol

Doesn't that suggest the training method is flawed?

For training, you need to follow a gradient, and usually the gradient is small enough that you need the precision. Training in 8-bit is fairly new but Google and now Grok seem to have gotten good results.

And Microsoft.

So these cards are geared towards inference ? There's a register article showing amds new instinct accelerators have much better.FP64

The article has numbers for both training, I think in FP8, and inference in FP4.

> next-gen NVLink switch that lets 576 GPUs talk to each other, with 1.8 terabytes per second of bidirectional bandwidth.

Well that's a pretty big step up from the 18 for the H100...

Explained by AI

https://xplained.vercel.app/ask?q=Nvidia%20Blackwell

Additional details on vRAM and spec: https://xplained.vercel.app/ask?q=Nvidia%20Blackwell%20vRAM%...

It's terrifying to think about how many watts of power are being spent on generating useless trash

This is a bit vague, are you talking about this new GPU/LLMs, social media, industrial society, or something else?

Most human consumption is useless trash by some metric.

Fun fact: each human uses an average 100W of power and generates mostly useless crap.

> each human uses an average 100W of power

Does that include heating, electricity and fuel spent for the benefit of the human? How about infrastructure? Externalities are huge.

Reply on news.ycombinator.com