Nvidia signs up for an Italian Job: Building for Europe the 'world's fastest AI supercomputer' by 2022

Europe is to build four Nvidia-Intel-powered supercomputers, one of which will be the most powerful super yet built for AI applications, the GPU giant reckons. That top-end machine, nicknamed Leonardo, is expected to reach 10 exaFLOPS albeit at FP16 precision; supercomputers tend to be benchmarked using FP64, though FP16 is …

[Claverhouse]

The Biggest and The Best

Whichever creep is president --- including any of the runners-up if either old fool conks it --- will immediately order a superior supercomputer.

[VexedGuru]

Re: The Biggest and The Best

And meanwhile in good old blighty .............the tumbleweed rolls silently across our bleak supercomputing landscape

[RM Myers]

Good old blighty

Have you just issued a call to Arm(s)?

[Vikingforties]

Re: Good old blighty

Nah, it'll cost an Arm and a leg.

[druck]

Re: Good old blighty

They'll need an arm to point all those FLOPs in the right direction.

[Elephantpm]

Re: The Biggest and The Best

Not exactly.

https://www.google.com/amp/s/www.marktechpost.com/2020/10/07/nvidia-announces-cambridge-1-uks-most-powerful-supercomputer-for-ai-healthcare-research/%3famp

[VexedGuru]

Re: The Biggest and The Best

Fair point and kudos to Nvidia and Jensen for stepping in to help rather than sitting on their hands, but the Italian machine is >20x the AI performance of Cambridge-1, I would respectfully suggest that that is not only "not in the same league" but a completely different sport.

[Gene Cash]

The question is no longer "will it run Crysis?" but "will it edit a Slow Mo Guys video?"

[devTrail]

Precision

That top-end machine, nicknamed Leonardo, is expected to reach 10 exaFLOPS albeit at FP16 precision

All four computers are capable of running simulations at higher and lower precisions, including 64FP and 32FP as well as bfloat16 and 8-bit integer.

I don't get it. Does it mean that the first sentence gives the number in FP16 precision just to get a awsome number, but then at runtime FP64 might be used? If that's the case, isn't that a scientific, EU funded, project? Shouldn't they avoid this kind of marketing tricks in this context?

[devTrail]

Precision for AI

though FP16 is presumably good enough for AI

FP16 might be good enough, but only for certain problems and only if the data normalisation and scaling is not too much.