Slashdot Mirror
Weta Digital's Render Farm Upgrade
Headspace2 writes "Weta Digital (The graphics company behing LOtR computer effects) has just purchased 220 2.2GHz dual Xenon machines, each with 4GB of ram, to add to their current render wall of 350 1 Ghz P3 systems. They have also placed an order for another 256 Xenon servers. And it's all running Linux. My favorite
quote is 'it is thought the server farm will be the most powerful processing site in the Southern Hemisphere'. They should use that in the FotR ad campaign... 'Rendered using the most powerful processing site in the southern hemisphere' Congrats the guys that get to play with all those clock cycles.
Make more movies.
Xenon is an element. Xeon is an expensive CPU. I see "Intel Xenon" too many times at work. Please not on Slashdot too.
High-speed Road Trip (18.000KPH)
Dual Xeon ?
$45 per U Colocation Special
How about XPs modded into MPs?
All from Gamepc but here they are
Pentium 4 Xeon Multi-Processor Under Linux
Prestonia Xeon 2.0 GHz vs. Athlon MP 1900+
kawai
Maybe you haven't noticed, but bus speeds have increased since you bought that Pentium 75 system. Though not as much as cpu speed, because that's historically been the focus of the, uh, personal computer. Who cares about optimizing network transactions on a PC? They were built to get away from mainframes, remember? Well, that was true 20 years ago and the paradigm has stuck for longer than it should have.
Even so, Consumer hard drives can now claim ATA-133 speeds, that's probably an order of magnitude faster than the 1.2 GB drives from five years ago. And SerialATA is coming. On the server side, I think U320 SCSI is out now. SCSI started at 5, now it's at 320. THat's like 64 times faster.
RAM has kept up, too. The first DIMMS were 66MHz, now you can get effectively 400MHZ DDR, or faster than that if you want soon-to-be-out-of-business RAMBUS.
Heck they invented the AGP port so we could play games, and that's at 4X now, with 8X on the horizon and some really bigtime advances in GPU power in just the past two years.
None of these have seen the speed increases of the CPU, but they are moving along at a nice clip. The PCI bus is maybe the weakest link here, but it's gotten better.
I think there's a lot of room for growth left in the current physical materials. I keep hearing 15 years until we hit the quantum barrier in CPUs, if we keep up with Moore's Law. There was a great article not so long ago about hard drives, and how they are basically doubling in areal storage density every year. In ten years, you can get a 120 Terabyte drive. Only one problem: What the hell would you put on it to fill it up?
Kinda like the predicament they find with broadband. There's nothing else to do with all that bandwidth than download mp3s and pr0n and warez. Oops.
Between 3 and 4M could include housing, space, support of various types, and whatever they payed for the logic of the switches. Further, the figure was a guess by outside sources.
:)
I'd say we don't have much to really go on. Esp. when the article says:
"Weta has been quiet on its use of technology after a deal with IBM this year was misreported around the world at around US$10 million. Sources now suggest it was more like a tenth of that price. "
Others have tried a price game and were not very precise
Umm, graphic (or video) cards don't actually do any rendering. It's always the cpu. The only thing you need a high-end video card for is (pre)visualization when you are modelling. No sir, it's just that roomful of noble gas doing the work there.
"During times of universal deceit, telling the truth becomes a revolutionary act." - George Orwell, 1984
There isn't...
b ooksea rch/isbnInquiry.asp?userid=66NT518KIO&isbn=0521635 039
Just think of it in terms of hilbert spaces (or just plain vector spaces). A qubit spans a 2 dimentional hilbert space. A (normalized) state on that space could be written
|S>=a1|0>+a2|1>
just think of |> as a vector, where a1 and a2 are ANY complex numbers such that |a1|^2+|a2|^2=1
for two qubits then you just have a 4 state space
|S>=a1|00>+a2|01>+...
for more info check out:
http://search.barnesandnoble.com/textbooks/
I looked on the Foundry website, 'only' 10Gbit.
I hate those exponential powers!
I am not a robot. I am a unicorn.
A blend of Maya, Softimage, Houdini, proprietary code, and Photorealistic Renderman
I agree with you, but to play Devil's Advocate, there are sometimes reasons you want a fast CPU, not just a fast cluster. Our SGI sales guy often tries to make this point, for the obvious reasons, but it's true.
If you are rendering out a large number of frames, you want the most possible aggregate CPU power, because rendering is extremely parallelisable (each frame stands alone). So 50 Athlons is better than 40 Xeons. But if you are just rendering out a 5- or 10-frame test sequence, and the wall is not already overloaded - then you want the 40 Xeons instead, since each one can take a frame and you'll get your result back faster.
There is also the issue of network bandwidth. In some cases you can benefit quite a bit by having fast boxes with as many CPUs per box as possible. This is because there is a non-trivial network burden in sending out the job to be rendered, along with all its textures, images, etc. This can be mitigated by multicasting and caching - but I don't know to what extent Renderman does this - but I know if you don't design it right, it can really slow down your jobs. (We evaluated one render distribution system that relied on Windows SMB file sharing for its I/O. Sending a 200-MB job to each of 10 render crunchers pretty much killed it.)
"How can you claim that you are anti-crack, while still writing a window manager?" — Metacity README
a cheap Mac Dual G4 1 Ghz is FASTER.
Its $2799 and comes with a 300 dollar DVD-r burner as part of the price.
It gets over TWICE as many RC5 keys per second than the fastest dual AMD MBs... and per dollar, the XEON is way slower than a Mac at RC5.
Macs have a huge L3 cache and AMDs have no l3 cache so that might be one reason macs are twice as fast, plus a cold page of memory can be simultaneously write and read faster on a mac than a pc typically. That helps for some benchmarks perhaps.
but if you go to TOP500.org they maintain a ranked FRESH list of all the top 500 render/gfop cpu farms.
xeons are rare in the list and Powerpc boards dominate it. check out the list yourself.
Admittedly a better list would be megabyte per gigaflop... a conttest each fall determines that and typically dual Pentium 3 boards with three netwrok cards each win that award.... not this overpriced xeon garbage.
I like xoen for one thing... it has PCI-X now and has for over a year and no AMD has PCI-X shipping yet. Check Pricewatch.com yourself.
PCI-X will ship on macs soon and tahts all I care about.
Just the CPU. You want good floating point support [which is why Titanic used 500 dual-Alpha boxes], and memory bandwidth, and of course lots of Hz are always nice.
Theoretically a renderer could use a GPU for a coprocessor, but I believe render software is so complex that any GPU on the market today would be too specialised to be of much use. Hardware acceleration works for games because the game developer can tailor the rendering requirements / algorithms to the capabilities of the hardware (as abstracted by OpenGL and Direct3D, or via vendor extensions to same). Render software, OTOH, is at the mercy of what the modeler / animator / compositor wants, and they are not willing to settle for "whatever the hardware can do the fastest".
"How can you claim that you are anti-crack, while still writing a window manager?" — Metacity README
They don't ALL work on the same frame at once. But it can be set up so that more than one machine works on the same frame: each renders a part of the same frame, then they are automatically pieced together by Alfred (the distributed rendering software that comes with Photorealistic RenderMan).
Massey university just announced that it is going to build a 128 node beowulf cluster (no imagination necessary!). Auckland University have recently got an IBM Regatta class machine.
Just a (quite impressive) stone's throw away from Weta is NIWA's Cray T3E
bash-2.03$ uname -a
sn6908 kupe 2.0.5.51 unicosmk CRAY T3E
I love running that uname :-)
Backups are for wimps. Real men post their data in comments and have slashdot mirror it