Rent A Bit Of Weta Digital

An anonymous reader writes linking to this story at stuff.co.nz, excerpting: "Five hundred powerful computers used by Weta Digital to help create the special effects for the Lord of the Rings may be put up for hire.... The pizza-box sized IBM blade servers each incorporate dual 2.8 gigahertz Intel Xeon processors and 6 [gigabytes?] of memory." Update: 03/22 07:08 GMT by S : The linked story says 6 megabytes of memory, we don't believe 'em.

Comments

Imagine

[Anonymous Coward]

A beowulf .. oh wait ...

Cost?

[nb caffeine]

What would this cost? Do they charge something like cpu/hours or the like? Will the average person have the ability to rent some clock cycles? I just want something that will be able to run doom3 when it comes out.

Pizza-delivery!

[The_Ace666]

Now where can I find a pizza-delivery company to get one of these babies delivered to my door?

Re:Pizza-delivery!

[DigiShaman]

Either way, they'll be hot! (pun intended)

Update

[hlopez]

Update: 03/22 07:08 GMT by S
-we don't believe 'YOU-

One thing to say...

[linuxkrn]

seti@home!

Re:One thing to say...

[CGP314]

folding@home

I used to run seti@home instead of folding@home, but then one day I realized I needed to switch. While finding extraterrestrial life would be the most important development in human history to date, the chances of finding it in my lifetime are very small.

On the other hand, the chances of my getting cancer or any of the other of the diseases folding@home works on is very great. Plus, if folding@home cures any of these diseases, it will extend my life and increase the chances that extraterrestrials will be found within my lifetime.


-Colin

Re:Distributed.net...

[Motherfucking Shit]

Imagine distributed.net being a CPU co-op. They take problems from clients in need of a ton of CPU, farm it out to distributed.net members, and at the end of the month/year you get a small check for all the CPU cycles you spent helping solve problems.

This was already tried, by a company called ProcessTree. The idea was that they'd sell your CPU cycles out and you'd get a cut. They also had it set up in a pyramid fashion, so that you also got an extra few cents for each person you referred to the program.

The best I could find was this mirror of the FAQ. Since ProcessTree.com now belongs to a domain poacher, I'm guessing they never did find a paying client...

Renderfarms online - old news

[Anonymous Coward]

Posting anon as I have an interest in some of these companies :
http://www.respower.com/ - 250+ machines (~500GHz), 250GB ram
http://www.rendercore.com/ - 700 machines
http://www.render-it.co.uk/ - 82 cpus (131GHz), 82GB ram)

The only 'interesting' thing here is that it's WETA's farm. Other than that, I doubt they offer the wide selection of software (lest they struck deals lately) not to mention field experience with 'oddball' files.

Good luck to them, though

LAN Connection ?

[ultranova]

Surely they used Token Ring to connect them ?

Nasty hobbits. They tricked us! Thieves!

[pariahdecss]

Nasty fat hobbit probably sold the extra RAM to buy Twinkies(R)

Re:I'd like to run ray tracing real time on this

[troon]

I'm rather tired of waiting for graphics to progress to the level they will be in in the year 2010 or so.

Just give it six years or so, and you should see the improvements you are waiting for.

Sell 'em on e-bay

[GloomE]

Wouldn't they make more by selling them as (framed) collector's items?
Blade 1 of 500: current bid $1(insert zeros here).

interconnect

[painehope]

the real killer is that there's quite a few industries that can't rent time on their cluster because the gigabit interconnect ( IBM blade chassis have a switch module internal to each chassis, and I don't think you can get any HSLL - high-speed, low-latency - network interconnect modules ( Myrinet, SCI, Quadrics, etc. ) for them ) has too high of a latency for their applications.

Bandwidth-wise they should be fine, as each chassis has at least four ports that could be trunked to a top-level switch w/ a beefy backplane ( I could tell you the # of ports per chassis if I was at work, as I've been messing w/ some of their blades lately ), giving a peak per-chassis bw of > 400 MB/sec.

Of course, I'm wondering how Weta got around it themselves, as I would think that rendering digital video is fairly heavy on inter-node communication. This would still be aswesome for web-servers or problems that are "embarassingly parallel".

Re:interconnect

[2megs]

Rendering digital video is about as parallel as compute loads get. Generally each frame can be an independent computation. For most ray-tracing algorithms, computing each pixel of each frame is fully parallelizable too.

The global AI things they did to have 10,000 troops all interacting together is obviously not quite so independent, but I'm willing to be the bulk of the compute load goes into creating pictures of those interactions, not the interactions themselves.

Re:Maybe they're right

[slash-tard]

Xeons only go up to 4 megs of cache and those were just recently released. At the time these were bought the max was 2 megs.

Re:I'd like to run ray tracing real time on this

[QuantumFTL]

Check out http://www.worley.com/fprime.html

My part-time employer (when I'm not working for NASA/JPL) Maas Digital just bought a copy of the software... it utilizes stochastic methods to allow flexible real-time raytrace rendering (with good motion blur!)

It turns out that motion blur in 3D graphics is a very hard problem because it's essentially a high-dimensional integral, and it turns out the best method of doing generalized high-dimensional numerical integration is a stochastic algorithm (monte carlo method) so it's not surprising to me that it's a great way to do motion blurs.

My favorite aspect of stochastic methods is their ability to be continuously refined (for instance, in a video game, the longer you spent looking at an object, the better it would get etc, and the graphics performance would degrade very smoothly with changes in system load etc). It is also ideal for parallel processing, as it can be dynamically parallelized to completely heterogeneous computing nodes.

Dan and I agree that there's going to be a lot of stochastic algorithms in the future of computer graphics (though he is hopeful that analytical methods will eventually make a comeback, as they have better asymptotic performance).

Cheers,
Justin Wick