Cray to launch new line of supercomputers

[kl0012]

... this time based on Westmere-EP, Nehalem-EX together with "Fermi" Tesla GPU.
http://www.theregister.co.uk/2010/03...idrange_super/

[SKYMTL]

They won't be including Fermi GPUs in these. Rather, it will be the current Tesla cards.

[OverShocked]

^ Thats strange...

[kl0012]

They won't be including Fermi GPUs in these. Rather, it will be the current Tesla cards.

Well, not yet, but it probably will be included after the official Tesla "Fermi" launch. Rather I was surprised that they didn't used ATI GPUs. After all Cray always has a tight friendship with AMD. Also some other vendors offers ATI GPUs in theirs server lines.

[natty]

Will they run Crysis?

[[XC] Oj101]

To me this isn't a Cray, anyone can do this from their backyard. This isn't a supercomputer, this is a massively parallelized Intel workstation.

Intel chipset? It's not theirs. Intel CPU? It's not theirs. DDR3? It's a standard. SSDs? A standard, and not made by them. Tesla GPUs? Not made by them. InfiniBand? It's a standard. Nothing here screams innovation like the old freon cooled Cray supercomputers.

[zalbard]

Nothing here screams innovation like the old freon cooled Cray supercomputers.

It's pretty hard to compete with Intel these days.

[Sam_oslo]

To me this isn't a Cray, anyone can do this from their backyard. This isn't a supercomputer, this is a massively parallelized Intel workstation.

Intel chipset? It's not theirs. Intel CPU? It's not theirs. DDR3? It's a standard. SSDs? A standard, and not made by them. Tesla GPUs? Not made by them. InfiniBand? It's a standard. Nothing here screams innovation like the old freon cooled Cray supercomputers.

This is how DELL, HP, Acer and others make their PC, too. Nobody needs to reinvent the wheel, but the innovation in this would the supercomputer in a little box.

[p2501]

To me this isn't a Cray, anyone can do this from their backyard. This isn't a supercomputer, this is a massively parallelized Intel workstation.

Intel chipset? It's not theirs. Intel CPU? It's not theirs. DDR3? It's a standard. SSDs? A standard, and not made by them. Tesla GPUs? Not made by them. InfiniBand? It's a standard. Nothing here screams innovation like the old freon cooled Cray supercomputers.

What a load of bull. Cray, like every other supercomputer maker on the market today, is about tightly integrating hw into complete systems, they never invented new cpus, memory or the likes.

They do, however, layout their own cpu/memory boards, and which is propably more important, they specialize in creating cpu interconnects for use on their boards (seastar).

And if you want to make this about phase change, it's still in their option list today.

[nn_step]

What a load of bull. Cray, like every other supercomputer maker on the market today, is about tightly integrating hw into complete systems, they never invented new cpus, memory or the likes.

They do, however, layout their own cpu/memory boards, and which is propably more important, they specialize in creating cpu interconnects for use on their boards (seastar).

And if you want to make this about phase change, it's still in their option list today.

Correction: They NO LONGER invent new cpus, memory, or the like.

Before Integrated circuits reached acceptable performance, it wasn't unusual for super computers to be built of hand picked transistors, capacitors, inductors, and resistors.


in fact it wasn't until 1974 that integrated circuits could surpass a set of hand picked discrete parts in clock speed

[Movieman]

So let me get this straight, esesentially they took a few dual 1366 boards, stuffed them in blades which meant they couldn't use the top 130w cpu's because blades suck at cooling, hooked up some fancy cables and some software and called it a supercomputer?
Did I read this right?

[trn]

These guys do know how to build a super computer if budgets do allow. Can't blame em for wanting to make a buck with some more economical gear.

I wonder if the kraken could break 100k WCG PPD?

[Movieman]

These guys do know how to build a super computer if budgets do allow. Can't blame em for wanting to make a buck with some more economical gear.

I wonder if the kraken could break 100,000,000 WCG PPD?

Fixed!

[JF-AMD]

Both Kraken and Jaguar are interesting computer. The blades are 8P, each consists of 4 2P nodes and a seastar ASIC for communication.

They did an upgrade of both systems last summer to 6-core from the old quad core barcelonas ( http://www.youtube.com/watch?v=ehCgG9yZQmg ) and this took them to the #1 and #2 on the supercomputer lists.

It was pretty cool to see the upgrade (I only saw the video) because they did it at the rate of ~5min/blade. The processing is amazing, but not nearly as amazing as the power and cooling complex that is needed because of the processing density.

I was out at CSCS, the swiss supercomputer center as they were unveiling their Monte Rosa cluster, which was also a Cray cluster. Very impressive. There is a picture floating around the web of me next to it.

[trn]

Fixed!

A quickie look at the numbers and I estimate that Jaguar listed at 1.76Pflops (not sure if this is up to date anymore) would produce 940,000,000 WCG PPD

*update I think Jaguar is listed at 2.33pflops after said upgrade. which would update that estimate to 1,240,000,000 WCG PPD, not too bad. If that joined team XS for a day we would crush easynews and team 2ch in a few hours... Anyone here a system admin in TN? (or have $65,000,000 to build your own?)

[>HyperlogiK<]

To me this isn't a Cray, anyone can do this from their backyard. This isn't a supercomputer, this is a massively parallelized Intel workstation.

So what is a supercomputer these days other than this? If machines based on custom vector processors running obscure proprietary instruction sets were what the market wanted then Cray would be selling them instead.

[[XC] Oj101]

What a load of bull. Cray, like every other supercomputer maker on the market today, is about tightly integrating hw into complete systems, they never invented new cpus, memory or the likes.

They do, however, layout their own cpu/memory boards, and which is propably more important, they specialize in creating cpu interconnects for use on their boards (seastar).

And if you want to make this about phase change, it's still in their option list today.

They absolutely used to. What they are doind now is using 90% of the shelf parts, man you can probaby order a modern Cray off of NewEgg if you're willing to build it yourself.

So what is a supercomputer these days other than this? If machines based on custom vector processors running obscure proprietary instruction sets were what the market wanted then Cray would be selling them instead.

Nothing, sadly. The supercomputer died to massive parallelization, and to me that is no longer Cray but more like a university project. For me it's not the concept but rather the name behind the project. Cray used to stand for so much more.

I just saw nn_step's post, that kind of makes mine redundant. The machine, even at $65m, can't compete with a $1k CPU under LN2 at superpi. Anyone can slap together one of these things whereas the early 70s were innovative.

---------------------------------------------------------------------
A typical discussion at Cray:
Engineer 1: Think of a REALLY HUGE number?
Engineer 2: 224,000!!!!!!!!!
Engineer 1: WOW! Lets build a supercomputer with that many CPUs
Engineer 2: HOLY %#&^ YEAH!! I'll call NewEgg so long...
Engineer 1: Awesome idea, I'll check Best Buy.
---------------------------------------------------------------------

[Particle]

Something being called a "supercomputer" is just a relative term indicative of its performance at the time of its construction. It doesn't much matter how you achieve it, be it 100% exotic or 100% mainstream.

[wuttz]

Both Kraken and Jaguar are interesting computer. The blades are 8P, each consists of 4 2P nodes and a seastar ASIC for communication.

They did an upgrade of both systems last summer to 6-core from the old quad core barcelonas ( http://www.youtube.com/watch?v=ehCgG9yZQmg ) and this took them to the #1 and #2 on the supercomputer lists.

It was pretty cool to see the upgrade (I only saw the video) because they did it at the rate of ~5min/blade. The processing is amazing, but not nearly as amazing as the power and cooling complex that is needed because of the processing density.

I was out at CSCS, the swiss supercomputer center as they were unveiling their Monte Rosa cluster, which was also a Cray cluster. Very impressive. There is a picture floating around the web of me next to it.

[wuttz]

umm sorry wrong pic..

[JF-AMD]

Does that supercomputer make me look fat?

And yeah, I dream of Opteron.

[p2501]

Correction: They NO LONGER invent new cpus, memory, or the like.

Before Integrated circuits reached acceptable performance, it wasn't unusual for super computers to be built of hand picked transistors, capacitors, inductors, and resistors.
[img][/img]

in fact it wasn't until 1974 that integrated circuits could surpass a set of hand picked discrete parts in clock speed

They absolutely used to. What they are doind now is using 90% of the shelf parts, man you can probaby order a modern Cray off of NewEgg if you're willing to build it yourself.

So you're both talking about a time when a real man was writing own drivers for his computer devices, wow. That is, in technological terms, so long ago it could be stoneage as well.

Makers of supercomputers do take ready made RAM, CPUs and other parts off the shelf for a long time now because it just isn't viable to be developing them by themselves anymore. That is true for Cray as for others, so I don't know where the thought "Cray isn't Cray anymore because they don't do everything by themselves" comes from.

They're about integrating standard parts in large scale and make it a efficiently working whole, which is difficult enough. And why would said maker of big irons voluntarily be missing out on the small scale systems, it helps them to survive and get money for, say, developing exascale systems.

[[XC] Oj101]

Please note that I said

To me

Cray used to be about innovative technological advances ahead of their time, whereas what they are doing now can be done with, relative to back then, incredible ease. Please don't quote me out of context by dropping the "to me"

[>HyperlogiK<]

Innovative technological advances are great when they do something which can't be achieved with more mundane existing technology, but in many cases the more mundane existing technology can be a better option. A good example was when AMD finally launched K10, it sucked. It might have been a more elegant and advanced design than Intel's MCM quads based on the NGMA but it was slower and had worse yields, thus it's elegance and sophistication was completely irrelevant.

[p2501]

Dropping the 'to me' part wasn't intentional.

But why do you think that only when they try to reinvent the wheel they are innovative enough?! You have Intels chips (differerent Xeon and Itanium versions), AMD's (Opterons), IBM (Power), collaborative things like Cell, Sparcs, MIPS and so on to opt between. Not to talk about GPGPU computing. And there's Nehalem-EX. Intel will, from what I gathered, only supply 4 socket chipsets (please correct me if that's wrong), if you want to get 8 or more to work together you'll be in for a round of inventing (like seastar+++++++ or something). Building a redundant, fault-tolerant, always online (99.9999% or more) large scale system from those components that doesn't choke on getting/saving data or melt itself, but rather works efficiently for the purpose it was build is nothing short of amazing.

To me, at least.