Intel-powered Aurora Supercomputer Ranks Fastest for AI

[unwind-protect]

It matters because it's not being done routinely. I know of only one modern production implementation and that's IBM Power10. Software-based emulation has heavy drawbacks, and would require extreme network performance in both bandwidth and latency.

It isn't just that. The bigger problem with virtual memory pages that are actually remote is that the local software can't tell which pages are fast and which ones are orders of magnitude slower. That is another reason why such schemes are unpopular. People rather use MPI if they can't have a single computer instead of mcgyvering together something with segfault handlers or userfaultfd.

 

[cvaldes]

Regardless of how people here are using the term "computer" or "cluster" it appears that the ISC conference attendees seem to accept the Intel system used for this benchmark as a "computer." It's not a question of semantics to those people.

But yes, there should be some sort of performance-per-watt metric included in these sort of assessments. It's not like electricity is free. As we know, performance-per-watt is a very important consideration for datacenter deployments. This is not a new concept.

 

[ncrs]

But yes, there should be some sort of performance-per-watt metric included in these sort of assessments. It's not like electricity is free. As we know, performance-per-watt is a very important consideration for datacenter deployments. This is not a new concept.

A press release was posted on TPU about the TOP500 rankings along with GREEN500 which is exactly what you want.

 

[Vya Domus]

I mean from a software/programming standpoint.

They can both be used in a unified fashion, it's just that one is purpose built to be better at it.

I can for instance write a program that multiplies very large matrices by distributing the load across multiple nodes on either of them and that program could also run on a single node. You could launch this program from just one command line terminal in the same way in either of those cases, it's just bizarre to me that you insist that a cluster isn't a computer based on a bunch of intricacies of how they work when it's obvious it's all just a matter of abstraction. One computer can in fact be comprised of several other computers.

It was about being a singular computer vs. a cluster of computers. Programming for a cluster is way harder than a singular computer, just like MT programming is more difficult than ST.

Of course it's different but you're still writing software for a computer. It's strange that you brought up the comparison with MT programming, do you think that when you write software that runs on one machine that this also doesn't count as a single computer because after all the processor could have many cores, so clearly since the software is adapted for MT then it can't be a single computer that you're using, correct ?

It's all just a mater of abstraction and scale, a single system could employ many other subsystems and you can use those systems to build an even large system.

 

[ncrs]

I can for instance write a program that multiplies very large matrices by distributing the load across multiple nodes on either of them and that program could also run on a single node. You could launch this program from just one command line terminal in the same way in either of those cases, it's just bizarre to me that you insist that a cluster isn't a computer based on a bunch of intricacies of how they work when it's obvious it's all just a matter of abstraction. One computer can in fact be comprised of several other computers.

Try it then, you'll see that it's not as easy as you claim it to be.

Of course it's different but you're still writing software for a computer.

So is programming on GPUs, and yet we drew clear distinctions for it.

It's strange that you brought up the comparison with MT programming, do you think that when you write software that runs on one machine that this also doesn't count as a single computer because after all the processor could have many cores, so clearly since the software is adapted for MT then it can't be a single computer that you're using, correct ?

No. Again, writing MT software in a uniform physical address space is a different paradigm from targeting a cluster of interconnected but disjoint (in terms of memory) computers.

This is my last reply to you because you're using the same troll-ish argument tactics as with our last discussion about SFUs in GPUs. You seem to have a Wikipedia-deep level of knowledge with an attitude that I have no more patience for.

 

[Vya Domus]

"It's more difficult to program", that's basically your argument at this point and you're telling me I am the one who is trolling here ? OK dude.

Fine, supercomputers aren't computers because they are clusters and clusters aren't computers because programming them is difficult and sharing memory is difficult and yet you could write software that runs on all of this. Now what's that software running on you ask ? Well, not a computer, experts say. Makes perfect sense.

 

[cvaldes]

Guys, sometimes words have multiple definitions depending on who is talking to whom.

As I pointed out the ISC conference people seem to be okay with Intel calling their thing a computer. Don't the people attending an HPC conference have some sort of idea about what a computer is? Do you see them bickering over it in headlines all over media? I don't.

Anyhow, sometimes TPU cuts-and-pastes corporate press releases so marketing lingo dominates. Get used to it because it's not stopping anytime soon.

 

[Vya Domus]

Don't the people attending an HPC conference have some sort of idea about what a computer is?

Obviously not, once again big W for forum dwellers. Intel, or anyone else for that matter, please take notes.

 

[cvaldes]

Cool.

So the people who design, build, purchase, configuration, operate, and program HPC computers don't know what a computer is according to some dilettantes on some anonymous Q&A forum.

Makes sense. Carry on.

 

[Daven]

Other than games...what is AMD faster than Intel at?

Everything per watt

You do realize that supercomputers are not intent for gaming, right?


He is focused on fabs and how many billions he can get to build them.
Intel stock up as $11B Apollo deal nears completion: WSJ
We can laugh at him constantly for the next 1-2-3 years, but if his plan works, if he builds those fabs, are competitive or even better than TSMCs, Intel will become again a successful behemoth and most of all, it's fate wouldn't be depended on x86 success/survival against ARM or whatever else comes in the future.

Isn’t that the case with all companies since the beginning of time…if they can?

 

[Scrizz]

I wonder how much longer Pat will last, he gotta be able to deliver things on time, or if they are late, blow expectations away. He's currently doing neither.

well... he wasn't with Intel in 2015....
He wasn't CEO until 2021, so can't blame him for Aurora