|
|
Log in / Subscribe / Register

LFCS: A $99 *not* supercomputer

LFCS: A $99 *not* supercomputer

Posted May 9, 2013 21:36 UTC (Thu) by daglwn (guest, #65432)
Parent article: LFCS: A $99 supercomputer

Ok, I work in HPC so I'm admittadly biased, but I also have deep knowledge of the domain.

This hardly qualifies as a supercomputer. It's cool technology, yes, and the price point is right. Massive parallelism *will* have to go mainstream, there's little doubt about that.

But a supercomputer is much more than a chip. It's a system: processor, memory, network, I/O, software. Each of these things is tightly coupled and highly tuned for maximum performance at reasonable (for HPC) prices.

So yes, hooray for Adapteva! But as far as I can tell Andreas never claimed Parallella is a supercomputer and neither should anyone else.

to post comments

LFCS: A $99 *not* supercomputer

Posted May 9, 2013 22:13 UTC (Thu) by jake (editor, #205) [Link] (3 responses)

> But as far as I can tell Andreas never claimed Parallella is a
> supercomputer and neither should anyone else.

Hmm, that's been said quite a few different places, by Jim Zemlin in the introduction of Andreas for one, as well as:

http://www.kickstarter.com/projects/adapteva/parallella-a...

which is the kickstarter, titled:

Parallella: A Supercomputer For Everyone

I can't say for sure that Andreas wrote that, but one would guess he could at least have put a stop to it if he wanted.

And the board does have:

> processor, memory, network, I/O, software

perhaps not in sufficient quantity/quality for your taste, however.

jake

LFCS: A $99 *not* supercomputer

Posted May 10, 2013 21:40 UTC (Fri) by daglwn (guest, #65432) [Link] (2 responses)

> I can't say for sure that Andreas wrote that, but one would guess he could
> at least have put a stop to it if he wanted.

It's too bad if Andreas is pushing this view. It's misleading.

> And the board does have:
> processor, memory, network, I/O, software

Well of course it does. It is a computer system after all.

> perhaps not in sufficient quantity/quality for your taste, however.

It's not about my taste at all. Ask anyone in HPC. The fact that processor, memory, network, I/O and software are robust, tightly coupled and highly tuned is what separates a supercomputer from a cluster and from Parallella.

Just as one example, GigE is nowhere near enough to handle climate simulations, modeling stars and analyzig combustion. It's not just about bandwidth and latency.

LFCS: A $99 *not* supercomputer

Posted May 11, 2013 13:26 UTC (Sat) by nix (subscriber, #2304) [Link] (1 responses)

Of course, there's no way that anything faster than GigE can be used in anything even remotely consumer-grade until it gets a bit cheaper. I considered going to 10GbE on the local net a while back, but looked at the prices, screamed, and decided maybe in ten years or when I win the lottery. Just a simple four-port switch was several thousand pounds!

LFCS: A $99 *not* supercomputer

Posted May 13, 2013 18:08 UTC (Mon) by daglwn (guest, #65432) [Link]

"It's not just about bandwidth and latency."

LFCS: A $99 *not* supercomputer

Posted May 11, 2013 3:42 UTC (Sat) by Trelane (subscriber, #56877) [Link] (2 responses)

It's not going to land on the top 500, no. I dont think that clustered supercomputers are the definition of supercomputer. Personally, I think that the high degree of parallelism warrants the use.

(I also work in HPC :-)

LFCS: A $99 *not* supercomputer

Posted May 13, 2013 18:10 UTC (Mon) by daglwn (guest, #65432) [Link]

> Personally, I think that the high degree of parallelism warrants the use.

Then a GPU board is also a supercomputer. Because that's basically what this thing appears to be.

But <fist bump> for HPC work. :)

LFCS: A $99 *not* supercomputer

Posted May 14, 2013 12:28 UTC (Tue) by jzbiciak (guest, #5246) [Link]

Of course, how are you going to feed a Parallella? Unless I misunderstood the diagram, the pipes into it are way too small. It'll work for exhaustive-search type algorithms with small working sets and high locality, but I'm skeptical it scales to larger problems well.

I work on processors that get pressed into HPC duty, and we see the rest of the system as being as important, or often even more important than the CPUs themselves. It's about machine balance.

(I thought about linking one of our latest chips, but then considered it might be bad form.)


Copyright © 2026, Eklektix, Inc.
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds