its a bargain!

February 27, 2005 6:35 PM

Cray Supercomputer for sale on Ebay. Starting price is only 10% of it's original cost!
posted by crunchland (39 comments total)

You're user #2975 and you post %#$% like this?

ps. This is not the first time a Cray has been on ebay, you know:
Results 1 - 100 of about 665 from ebay.com for cray computer.
posted by Mean Mr. Bucket at 6:47 PM on February 27, 2005

Sorry.
posted by Mean Mr. Bucket at 6:52 PM on February 27, 2005

I work with a lot of the guys who designed that sucker.
posted by substrate at 6:57 PM on February 27, 2005

You're mean, Mr. Bucket.
posted by ChrisTN at 6:59 PM on February 27, 2005

Mr.Bucket is on a roll tonight. Motherland can sleep peacefully.
posted by c13 at 7:05 PM on February 27, 2005

<3 not h8
posted by TwelveTwo at 7:13 PM on February 27, 2005

heartiez not hatiez

72ghz!
posted by angry modem at 7:22 PM on February 27, 2005

you post %#$% like this? ... at least he isn't mean.
posted by R. Mutt at 7:29 PM on February 27, 2005

NEW IN BOX!!!!!!
A++++++++++++++++

posted by scarabic at 7:53 PM on February 27, 2005

Mr. Bucket, don't make me put my balls in your head.
posted by cloeburner at 7:54 PM on February 27, 2005

wow. I had no idea it was so mundane to sell a cray on ebay. My mistake for taking some other bloggers word for it.
posted by crunchland at 7:59 PM on February 27, 2005

If in six years it is worth %10, when does it become effectively worthless (ie. the power and space to run it) - the add even says it has a "few years left". Seems like a very specialized market, one that needs a cheap Cray right now today for a short period of time.
posted by stbalbach at 7:59 PM on February 27, 2005

It would kick ass to have something like that in your apartment. Imagine having people come in, and a whole wall of your apartment is covered with a computer. How Star Wars.
posted by interrobang at 8:04 PM on February 27, 2005

Mean Mr. Bucket might be mean, but he's also a moron. Most of those 665 listings that he got his panties in a bunch over are things that aren't Cray supercomputers. In fact if you search for Cray in the computer section all you come up with is the linked T3E and a few bits and pieces.
posted by substrate at 8:06 PM on February 27, 2005

If you sign up for PayPal Buyer Credit, you get $10 off!
posted by Johnny Assay at 8:18 PM on February 27, 2005

Damn, you just know that some mad scientist somewhere is rubbing his hands together with glee over this auction.

Until it can run Unreal Tournament 2004, I'm just not interested.
posted by fenriq at 8:32 PM on February 27, 2005

What I'd like to see is not UT 2k4, but rather Doom 3 *RAYTRACED* at 1600x1200 @ 85fps. I imagine this kind of rig could pull that off, because raytracing lends itself to a fair amount of parallelization.
posted by Ryvar at 8:46 PM on February 27, 2005

So... there are no bids. Is this thing worth several hundred thousand dollars to anyone, anywhere? I'm not geeky enough to understand if this thing is still powerful enough to blow past a brand new machine costing 1% as much, but it seems unlikely. Anyone?
posted by scarabic at 9:04 PM on February 27, 2005

Nice racing stripe.
posted by wfrgms at 9:07 PM on February 27, 2005

It's too bad that it has an automatic low bid. It might've actually been interesting to see how much it would go for. "I hear $50 from the nerd in Minnesota, do I heard $75..."
posted by Arch Stanton at 9:27 PM on February 27, 2005

Weren't Crays used for nuke design? I wonder what would happen if someone from Iran of NK placed a bid.
posted by c13 at 9:29 PM on February 27, 2005

It has 120 600 Mhz Alpha processors in it, of which 112 are actually available to run user code. Those really aren't that fast anymore.

Using Opterons, you could build a 128-proc cluster for, oh, $175,000-$200,000, including network hardware. For many applications, the A64 cluster would be faster. The Alphas were good chips, but these are only 600Mhz, and a 2 gigahertz Opteron would kick butt in comparison.

The data busses betwen processors would be much slower, so the Cray would still win in some circumstances. Usually, though, when you need this many processors, your problem doesn't require that much communication between them... so most of the time the much faster A64 execution engine would trump the Cray's better internal bandwidth.

Also note that the memory and storage are relatively wimpy... 30 gigs and 1.2 terabytes, respectively. That's really tiny by 2004 big iron standards.

I can't really see this one selling. Much like a desktop from 5 years ago, it has very little resale value. One potential use would be in a shop that already had a similar Cray and needed to increase their processing power a great deal in a short time. This model will go to 2,048 processors, but you probably still pay a mint for those... it might be cheaper to buy a used system and expand that way.

It could also be useful for parts, but 350k for parts... ouch!
posted by Malor at 9:30 PM on February 27, 2005

scarabic: This machine is still useful, but I don't know if it's worth 350 or 400k. Crays have always had high bandwidth memory and i/o compared to "normal" servers, so perhaps it's hard to compare it to off-the-shelf hardware. This one has 30GB of ram and 134Gflops. There are Opteron boards that support 32GB of ram, but you're going to need a bunch of cpus since an 850 (2.4GHz) does 4 gflops IIRC.

If you're building a cluster with several nodes, you probably get a price break from the manufacturer, but 8 machines with 4 850s and 4GB of ram each is going to be roughly:

850 cpu x 4 -- $4500
4GB ram -- $500
motherboard -- $500
case/power -- $500
1TB disk -- $1000

= $7k x 8 = $56k

On preview, add me to the list of people wondering whether this machine can be used for nuclear weapons design. My gut says that it can, but I have no experience in the field.
posted by beaverd at 9:42 PM on February 27, 2005

Nuclear weapon design was done -- in the 1940s -- on mechanical adding machines and IBM/Hollerith punch card rigs. A T3E is overkill for your average prenuke state.

In any case, $350k is a bit high, but if you have a line on additional PE capacity it's not utterly unreasonable. Access to the Cray toolset and very fast memory paths is worth quite a bit for certain classes of problem.
posted by majick at 10:13 PM on February 27, 2005

The problem is, at least for a prenuke state, is that this research is classified. So they'd have to do it themselves.
But, from reading this thread, they probably could do it a lot cheaper with modern components.
posted by c13 at 10:31 PM on February 27, 2005

I sincerely hope someone ends up buying this although I'd agree the asking price is perhaps a tad steep. As I understand it Cray has (or at least, had) a stringent policy of decommissioning machines themselves. Because of this it is very rare to see complete machines, let alone complete machines with the software and developer's tools. As for why no one has bid, I would wager that there are few people/firms that would know what to do with it. The machines are rare and people that know how to use them probably more so. Clustering commodity hardware is the more familiar path for many to take although the end result is probably far less elegant.
posted by darksquirrel at 10:41 PM on February 27, 2005

I second b1tr0t. A Cray 1 wouldn't even have to work to be worth quite a bit. I'd love one in the lounge room.
posted by krisjohn at 12:45 AM on February 28, 2005

This Cray won't make the top500 anymore. The fully upgraded version goes in at number 280.
In 1999 it was 67th. Just about what the Air force used to forecast the weather (or check their email, it doesn't really say ... ).

I like the way the eBay description says you can upgrade from 120 to 2048 processors, as if you can just buy the processors in PC World. Assuming you have $50 million to spend. That and the 1.3 trillion byte hard disk.

The seller only has 16 transactions, one of which is for a 97lb document scanner which can handle letter paper at 300dpi. I'd probably go with the escrow.
posted by Maxwell at 12:54 AM on February 28, 2005

Malor, you could build a beowulf cluster of opterons but there are a lot of real tasks that the T3E will still be faster on. The interconnect on the T3E is much faster than gigabit ethernet or networking over firewire in terms of latency, actual bandwidth and reliability. If the programming model you have is each processor grinding away and only occasionally needing to transfer data between other processors than commodity computing will win. If you're modelling something that requires a lot of communication between processors the T3E will still spank you slow processors and all.
posted by substrate at 5:59 AM on February 28, 2005

I wonder if Pixar could use it.
posted by crunchland at 7:08 AM on February 28, 2005

Interestingly the UK Met Office was using two T3E's as their primary machines for forecasting until 2002, when they replaced them with newer NEC SX-6's.
posted by rolypolyman at 7:09 AM on February 28, 2005

interrobang: It would kick ass to have something like that in your apartment. Imagine having people come in, and a whole wall of your apartment is covered with a computer. How Star Wars.

You would think, but it gets old really fast especially if your actually using the thing. Power requirements alone are rather impressive. I had a mini for a while and the second time I moved it I vowed never to move it again.

Pixar is probably better served by a massive cluster. Each frame render is a discrete job that doesn't require any communication with other nodes.
posted by Mitheral at 7:18 AM on February 28, 2005

c13--While you're worrying about rogue nations buying Crays to design nukes, some people worry about them buying Playstations.

As much as a Cray selling (or not selling) for $350K is an eye-opener, I think that is too.
posted by adamrice at 8:13 AM on February 28, 2005

interrobang: It would kick ass to have something like that in your apartment. Imagine having people come in, and a whole wall of your apartment is covered with a computer. How Star Wars.


I used to work for Convex Supercomputer. Down in manufacturing, when they would test boards, or obsolete boards, they would put the bad boards in a pile for rework/retest, etc. When HP bought Convex, they told us to dump all of the rework pile. Since they were going to be trashed, I asked for permission to take some of the motherboards...because I had a groovy sculpture idea.

I never did get the sculpture figured out, because the systems run so hot they need to be water cooled and my circuit board skills were not nearly as l33t as I'd thought, but I am the proud owner of supercomputer motherboards that are bigger than many side tables. These things are massive, loaded with processors and other chips and have the coolest looking circuit pattern.

Almost 10 years later, and I still haven't found a project for them...but they still make the geek inside me giggle with mad-scientist glee when I see them.
posted by dejah420 at 8:27 AM on February 28, 2005

I remember when I was a college student and I went on a tour of Lawrence Livermore labs. We were led through this incredible room full of Cray computers, and it really looked like science fiction with the water cooled processors displayed in some clear glass case with water just coursing through, it was amazing.
Talking to the people who ran the lab, they said the problem with Crays, wasn't that they weren't fast enough, it was that they didn't have software written that could fully exploit massively parallel architectures like Cray. Of course this was in 1993, so I'm sure much has changed.
posted by patrickje at 10:34 AM on February 28, 2005

The DOE's use of supercomputers (to test the effectiveness of aging stockpiles as well as testing the effectiveness of new designs) requires vast amounts of data and models that were developed through decades of live nuclear tests.

Even if Iran or N. Korea got their hands on a top of the line supercomputer, they couldn't get their hands on those nuclear models, which are some of our most closely guarded secrets.

In 1996 the French decided to perform additional Nuclear tests. The purpose of these tests was to perfect their models, though we had offered them our own.
posted by exhilaration at 12:57 PM on February 28, 2005

substrate, if you'll notice, I did say almost exactly that -- that for some classes of problems, it would still be much faster than commodity hardware. It has a link roughly comparable to HyperTransport on every processor, which is obviously vastly faster than gigabit ethernet.

Admittedly, this isn't a terribly educated opinion, but it seems like most of the programs that parallelize well don't seem to need that much crosstalk. That may be an artifact of people rewriting algorithms to not need crosstalk precisely BECAUSE it's expensive.... but regardless, it seems like a large number of modern problem types are fine on the 'fast proc/slow interconnect' Beowulf model.

I'd bet my boots, though, that nuke sims would requre boatloads of crosstalk. Everything interacts with everything. For that kind of sim, this Cray would probably be far better than anything you could do with PC-type hardware at the same price point.

However, that said... remember that they were talking about labeling PS2s as 'weapons-grade processors' and refusing to allow Sony to export them? From that, I believe that networked PCs are probably more than adequate to write nuke sims. This Cray will almost certainly be better in terms of bang-per-buck, assuming that this base model is enough for the simulation in question. Off-the-shelf stuff, while probably more expensive to reach this level of performance in this specific problem, would both start cheaper and scale much higher. Expanding the Cray would be monumentally expensive. Adding PCs to a cluster is not.

And for many classes of problems, the PCs would be much better.
posted by Malor at 1:41 PM on February 28, 2005

What about AI applications? Neural network? Feasible?
posted by squidlarkin at 2:19 PM on February 28, 2005

It's unclear to me why the government bothers to restrict the export of powerful computing technology. Building a simple nuclear weapon is mostly a problem of figuring out how to refine enough ore into material of high enough purity that it will DTRT. Large/fast computers are of no particular help.

Simulations are only required to ensure that complicated weapons will work, especially if you don't want to actually set many (any) off.

Simple nuclear weapons are quite sufficient to scare the pants off most people.
posted by nickp at 11:30 PM on March 1, 2005