Red Storm Rising: Cray Wins Sandia Contract

anzha writes "It seems Cray is alive and kicking at least. They might even be making a come back after its very rough time as a part of SGI. The big news? Cray seems to have won the Red Storm contract - Sandia's newest supercomputer procurement - from Sandia National Labs. Check out the press release here. I'd say that this is probably an SV2, but the press release is a bit scant on details."

Comments

Care to shed some light?

[korpiq]

I know something about the problems of multiprocessing, but I would like to know how come monumental systems can still sell in the days of commodity hardware and - oh gosh, not again - Beowulf :). Ok, maybe spread-out computing is not applicable to all kinds of computationally heavy problems, but I'd really like to see some stats on where a monolith like Cray is more applicable and where a multilith like what they use in movie rendering.

Just pondering while waiting for net-enabled market of processing power (remember processtree?) and storage space (freenet, the new one) to make millionaires of all excessive-hardware owners, through paypal. Well, maybe not :)

Re:Care to shed some light?

[UranusReallyHertz]

It all comes doen to inter-cpu bandwidth needs for the particular piece of code you need to run. A render farm has pretty much no need for interprocessor bandwidth, wheras crays have it in the 100's of GBs/s because the kind of numerical physics simulation that is usually run on these beasties needs all the bandwidth it can get, and a little beowulf cluster of x86 toys just ain't gonna gut it.

Re:Care to shed some light?

[bentini]

A lot of problems just don't scale well across different systems. Say, fo rinstance, you need to simulate particle interaction...

Distributed processing is only really good when the subproblems are separate enough that they can be calculated separately.

Also, supercomputers are a lot better for vector code. Intel and Athlon might say that their current offerings are Vector Processors, but they really aren't. When you need to exploit DLP, supercomputers are the way to go.

Also, research and funding like this will uncover the techniques that we can expect to be exploited in desktop processors in 5-20 years, so it helps us eventually.

Re:Care to shed some light?

[korpiq]

Ok, I'm only halfway through the video about SV2 architecture at http://www.cray.com/company/video/ and I already find my question laughable :D

This stuff is plain cool.

Re:Care to shed some light?

[4of12]

I'm not sure what this new Red Storm machine has in the way of individual nodes, but Sandia has some history in parallel computing, dating back to the paper

Gustafson, J.L., G.R. Montry and R.E. Benner. "Development of Parallel Methods for a 1024-processor Hypercube." SIAM Journal on Scientific and Statistical Computing Vol. 9, No. 4, July 1988.

as well as the ASCI Red machine, which, IIRC, was the first machine to break the 1 teraFLOPS barrier.

That machine, BTW, was built by Intel out of fastest Pentium chips of the day. I think a later upgrade to Pentium IIs increased its speed to about 3 teraFLOPS.

As far as MP machines are concerned, it could be argued on the basis of the ASCI Red machine that they have a fairly "economical" strategy [I know, I know, it's hard to argue that anything costing $9e7 as being "economical" - but you are talking about buying one of the fastest few computers in the world - rack mounted Athlon MPs could do great until you get up to O(100) processors, but doing the interconnects for O(10000) processors gets to be tricky].

Also there is CPlant, their own (everybody's gotta have one) pet project to build a B----- cluster out of Alpha based machines running a modified Linux.

generally

[martissimo]

Crays run a variant of Unix called UNICOS

too lazy to dig you up a link, and no need to karma whore anyways, just google it or go to cray.com and read about it ;)

Associated clip from the industry standard

[jukal]

from: Building a Better Bio-Supercomputer, this one year old newspiece might provide some info on what the system will be:

<clip> Competitor Compaq is taking a different path. In January, the company announced plans to develop a 100-teraflop bio-supercomputer dubbed Red Storm in partnership with Celera Genomics, the Rockville, Md., company that mapped the human genome, and Sandia National Laboratories in Albuquerque, N.M. Although Blue Gene will be 10 times faster than Red Storm, a Celera executive stresses that the company's machine could eventually match IBM's speed. Unlike Blue Gene, though, Red Storm is being designed for a broader array of life-science experiments and may be used to conduct nuclear research. The supercomputer, set to begin operating in 2004, will cost an estimated $125 million to $150 million to build. </clip>

This seems to be somewhat in-line with the cost approximate stated in the press release $90 million. Or am I completely in my effort to undestand what this press release is about?

Re:I wonder..., NO, Crays (SV1, 1+ and 2) use CMOS

[anon mouse-cow-aard]

The answer to the question is no.. Cray doesn't use ECL for the main beasts any more. That was one of the things that drove them into the ground in the 90's. The Japanese switched to CMOS, and drove the prices way down. Cray eventually followed suit, with their former low-end (YMP-EL which was CMOS based from the get go) spawning the SV1.

More Info and doh!

[anzha]

Since I submitted this story, Sandia National Labs has released their own press release here. Note that they say, it's an MPP (Massively Parallel Processor), but details to come.

What's interesing is that Cray has two machines that might be called MPPs:

1. The T3E with it's single system image, Unicos/mk and Alpha processors.
2. The Linux Cluster.

The SV2 might be called a massively parallel vector machine with potentially thousands of vector processors; However, they likely would have said 'vector' in the initial press release. On top of that, Cray would have trumpeted probably quite loudly they'd sold $90 million worth of SV2 because it helps more systems.. That makes me have doubts whether or not its an SV2.

The MTA doesn't count here either being called a multithreaded architecture rather than a parallel one (semantic hair splitting, yes, but important ones).

Furthermore, Cray is in the process of discontinuing the T3E because of its age.

To make it even more delicious is that Red Storm is mentioned a lot in searches at Sandia in conjunction with Cplant. Cplant uses linux...

So with a little bit of thought that would imply which Cray would be used here?

Saying 'imagine a beowulf cluster of those' might be a bit more accurate than the joke would normally go. ;) BTW, sorry, I can't believe I missed the w. Is Bush holding it hostage in his name? ;)

supercomputer "corporate welfare"

[peter303]

Generally the only way large supercomputers can be built in the USA is through government contracts. Industry is unwilling to pay more than $10 million for a large computer. The Department of Energy has been commisioning top-end computers ($10M to $100M) for weapons research and NOAA for weather forecasting.

I am ambivalant about this. On one hand I want to see a petaflop computer by 2010. (Two 100 teraflop computers have contracted for the 2007 timeframe, so this is possible.) On the other hand I am suspicious that computer companies won't build these on their own and dont like the governement propping up weak computer companies.

Re:Huh?

[fgodfrey]

I realize that this is a troll, but I'll reply anyway. The technologies that are invented for super-computing class systems have a way of making their way over time into a machine that you *will* put under your desk. Both MMX and more Altivec (a beefier version of MMX that is in certain PowerPC processors including the ones in I think all current Macs) are based on vector processing technology and ideas originally designed by Seymore Cray in the 70's for the Cray I. So the answer is, everyone cares, even if nobody knows exactly why. I won't pretend that I know exactly what technology from current Crays will end up in desktop systems in 10 years, but *something* probably will. The same can be said for the technology that is in current SGI and IBM (etc) supercomputers.

For another example, clustering technology (which I'm sure is going to get posted about in this thread) was an attempt to duplicate and borrows ideas from the massively parallel machines like the Cray T3E, the SGI Origin, and the old Thinking Machines boxes.