Interesting read on the intentions of ARM in the HPC space. Could we be seeing an ARM based system in the Top 100 at any point in the near future?


http://insidehpc.com/2010/09/13/arm-aim ... pc-market/

Why not, what ARM cannot do -now- with plain brute-force performance, they can do with quantity and better power efficiency. I.e. a system that performs the same FLOPS as a x86/x86_64 one, would definitely include many more ARM cpus, but OTOH it would waste only a fraction of the energy required.

Gross arithmetic: a cluster of 10 Cortex-A8 (eg. EfikaMX) at 800Mhz performs like 10 Pentium III systems at 1Ghz, more or less. But the efficiency is astounding, as each EfikaMX wastes ~7Wts while we know that PIII-class CPUs need more than 20 times as much energy. With today's CPUs we might compare to Atom, but in essence, today's Atom CPUs are actually slower -but at a higher frequency and they include better power management.

I think the next years we will be watching an interesting battle unfold between ARM and Intel :)

Konstantinos,

I understand your working on the NEON optimisations. Are there NEON optimised versions of ATLAS or other mathematical libraries available at this time?

Once I get my smartbook (hopefully this week) I plan to do the inaugural run of linpack...for the sake of curiosity.

None that I know of, but of course someone might be working on it. :-/

markos:

"I.e. a system that performs the same FLOPS as a x86/x86_64 one,
would definitely include many more ARM cpus, but OTOH it would
waste only a fraction of the energy required."

Consumed power during runtime is nothing compared to what is
wasted during manufacturing. Therefore all these CPUs should have
a much higher carbon footprint than the x86's.

The x86's won't stall in development if you think that? In five
years they may still be ahead of ARM..

You cannot make such simple claims as 'multiple ARMs vs. one x86' when it comes to cost of manufacturing.

There are things like system TDP requirements - x86 systems cannot be build as cheaply as ARM ones, as the former have much higher TDP requirements.

Then comes cost of ownership, which I don't think is negligible. Over an average lifespan of 5 years, an ARM system would so much outperform an x86 one in terms of power cost of ownership, that it's not even funny. I really can't fathom where that 'lower carbon footprint for x86' originates from.
I does not matter. Intel's lithography lead will not last forever (~5 years is what I give them, actually, before the industry globally hits the 'bottom' nodes). Once Intel lose that advantage, simple things like fundamental architecture efficiency will become entirely dominant for the silicon's power efficiency (much more so than they are today). The future is not rosy for Intel there.

Eh? That is probably false. I don't know the daily yields of cpu factories, but I'd guess it would be at least in the hundreds of CPUs/day (probably in the tens of thousands, but let's play safe). Assuming that a factory wastes a few hundreds of KWts of power(probably MWts but again let's play safe), that means a new cpu "costed" a few KWts of power to produce -I did try to find numbers for these, but Google wasn't much help :-/

However, assuming people use their computers for years -esp in a cluster that we're talking about- for 24/7, let's say 2 years before an upgrade, I think it's pointless to compare. An ARM cpu running at 7Wts/h, would definitely consume much more than a few KWts in the two years of its runtime (~122KWts to be exact). Compare that to 80Wts of a modern Intel CPU, which amounts to 1.4MWts in 2 years total).
Maybe, maybe not. Considering their later pace, I think ARM beats Intel in terms of progress rate -not absolute terms. What I mean is that ARM cpus progressed in a faster rate than it took Intel to achieve the same performance.

It will definitely be an interesting race, that's for sure. :)

Guys, sorry to spoil the party, but Intel is way more powerful than ARM. I'm not talking about the speed or dissipation in one of those products.

But as a company.

What stops them buying ARM outright, if they wanted or needed to?

You mean as a hostile takeover? Well, I guess that largely depends on who holds the control packages of ARM, assuming there's one, and whether those parties would sell.
Also, there are such things a anti-monopoly laws, which Intel has tasted in the past, IIRC. So I doubt they would do somehting as blunt as a hostile takeover to ARM.