Nvidia turns up the AI heat with 1,200W Blackwell GPUs

For all the saber-rattling from AMD and Intel, Nvidia remains, without question, the dominant provider of AI infrastructure. With today's debut of the Blackwell GPU architecture during CEO Jensen Huang's GTC keynote, it aims to extend its technical lead – in both performance and power consumption. Given Nvidia's rapid rise in …

[aerogems]

Who needs central heating when you can just play a few rounds of the latest FPS and heat your place with the waste heat? Shit, you could just pump it into the clothes dryer too.

What I want to see, is sort of where CPUs have been going. Improving the performance per watt ratio. If I could get RTX 20XX performance out of the power demands of an old GTX 9XX series card, great. Then try to improve it to 30XX and so on; 1.2MW just for the GPU is well past ridiculous. I just picture that scene from National Lampoon's Christmas Vacation where they turn on the Christmas lights and then cut to the power meter spinning like crazy.

https://youtu.be/oHVG2UrWNh0?si=hP4QD0VTSC8j6zLY&t=75

[cyberdemon]

The H100 was how many Watts? At least 600? If this really has 5x its performance then they -are- improving performance per watt?

Also, 1200W is 1.2kW, not MW ...

You can remove the tracking cookie off your u-bend links, too.

[aerogems]

Not sure what comment you meant to respond to, but it couldn't have been mine unless you willfully misinterpreted what I wrote. I know subtlety on the Internet tends to get lost, but I went out of my way to give a rather specific example. Maybe that's why you put a question mark instead of a period, you weren't sure you were responding to the correct comment.

Instead of increasing the amount of power drawn, work on improving performance within existing power consumption levels. Like how ARM chips can give performance within spitting distance of x86, but at maybe 20% of the power draw. The point at which video cards started needing their own dedicated power feeds from computer PSUs, that should have been the point at which either they worked to up the amount of power that can be supplied by the bus, or started working on efficiency rather than performance.

Though, you are right, I added an extra zero onto the end of things, so it is KW, not MW.

[David Hicklin]

>> I added an extra zero onto the end of things, so it is KW, not MW.

Stick a 1000 of these into your AI DC and you will have your MW.

These things are just becoming insane

[anothercynic]

Regarding the waste heat, I had the same thought - Just pump this through your central heating... it'll probably cost less than a heat pump ;-)

[katrinab]

It is a 1.2kW resistive heater. It will cost the same to run as any other 1.2kW resistive heater you can buy for a few £/$/whatever.

[Geoff Johnson]

1.2kw of heat, and anything else it does is a bonus. Are there any distributed cloud businesses that'll pay you for the use of it? Maybe resort to that old favorite, Bitcoin mining.

[Anonymous Coward]

But will it run Doom?

[Steve K]

Can it write Doom/Crysis

[katrinab]

Doom is entirely CPU I think? So probably not. Or at least not particularly well.

[Geoff Johnson]

Doom was all CPU, but it ran on a 386 at a pinch. You could probably simulate a 386 at the transistor level on this beast and run doom on there.

[HuBo]

Of tribal arts

I just love the "square robot face" aesthetics of the GB200, as shown in the (unfortunately tilted) 2nd image of the article -- one can clearly discern the two thick eyebrows, two beautiful square eyes, inspirational large square nose, and broad row of seven teeth, of the anthropomorph! It should really be designed for vertical placement on a gallery wall (or museum) IMHO (not hidden in some 1U cabinet drawer where no one will ever see it!)!

[HuBo]

Re: Of tribal arts

^{(pretending for a moment not to be myself, for sanity-preservation ...)}

Oh yes! add a 4U air-cooling tower to that Grace nose, a pair of liquid cooling shades to these Blackwell eyes, a potential goatie of PCIe connectors, or a fang of cards, to taste, and a spiky mohawk of colorful cables at the NVL-C2C junctions, and, WhAmO! A perfectly true-to-life digital twin of the GBH-200, hardcorest of the hardcores, fastest of the fasts! (borrow Jensen's leather jacket for extra effect!).

Stunning!

[IGotOut]

I'm getting some of these...

Just waiting for the nuclear power plant to be finished in the shed and then I'm good to go.

[Herring`]

Blackwell?

Is it named after Nigel? I'll get my DPAK.

(I did once ask ChatGPT to write a set of HMHB lyrics and the results were predictably shit.)

[Yet Another Anonymous coward]

Re: Blackwell?

David Harold Blackwell - American mathmo

[Steve Channell]

FP4 WTF

The distinction between G-Force and Quadro GPU used to be the latter sported full FP64 double precision floating point numbers while G-Force only supported FP32 single precision floating point numbers that were fine for games and mundane traditional calculations like Ballistic range calculation and Hydrogen Bombs

FP16 and later FP8 where billed as the advance that set Google Tensor Processing Unit (TPU) apart from GPU that were thought to be useful only for training.

With nVidia support for FP8 and FP4 in microcode, it's difficult not to see the end of any argument for dedicated TPU processors.