LinkBack URL
About LinkBacksany reason people aren't considering a C2Q with an older g35 mobo, would bring the build cost down further.
any reason people aren't considering a C2Q with an older g35 mobo, would bring the build cost down further.
How so? AMD boards are cheaper and the chips are about the same, AMD are generally regarded in the reviews shown to provide better rendering results.
http://www.nvidia.com/object/cuda_home.htmlOriginally Posted by Fraz
CUDA could be used for this, in fact thats what it is designed for. What does your brothers company use for rendering? If it is able to make use of CUDA there may be some test to show how performance compaires between a few graphics cards and a few processors.
You never know, perhaps a Tesla would be better bang for buck than a bunch of seperate PCs, especially when you take operating costs (electricity) and the size of it into account.
What about the hard drives? it doesnt matter if you have 1 million pcs running at 1 million GHz each if the bandwidth to the hard drives arnt up to the job you will run into the same bottleneck each time. I would look at hardware raid setups.
The first thing I thought when I read this thread was parallel GPU computing. As Funkstar suggested, CUDA could be the best solution. Do away with the cluster and instead, a single machine with several GPUs. One similar project, the FASTRA.
In which case, I suggest a 4x PCI-E motherboard and 4x GTX 295's
That isn't really a problem with 3D animation. Each frame could take a minute or even an hour to render, but it's going to be at 4K resolution at the very most (for cinema productions) which is the equivelant to about 8mp. 3D footage is generally rendered out to individual graphics files (TIFF and TGA are the most common) then combined into a movie using something else. This allows different parts of the render farm to render different frames and they can be combines easily. You could probalby render to a central server of NAS box and not loose much performance.
Just spotted that Fraz did say 3D Studio Max, had a quick look at the Autodesk website and it doesn't appear to be able to use CUDA for rendering, though it can make use of a GPU for previews etc. (http://usa.autodesk.com/adsk/servlet...12&id=10032262). Nvidia does own Mental Images so I would be surprised if there wan't a version at some point that supports GPU acceleration.
Last edited by Funkstar; 09-02-2009 at 05:45 PM.
Ok, so as an aside, I've just been to a 3 hour seminar on CUDA by Nvidia people, and I gotta say that saying "CUDA could be the solution" is a bit... well... vague/irrelevant.The first thing I thought when I read this thread was parallel GPU computing. As Funkstar suggested, CUDA could be the best solution. Do away with the cluster and instead, a single machine with several GPUs. One similar project, the FASTRA.
In which case, I suggest a 4x PCI-E motherboard and 4x GTX 295's
Yes, CUDA could well be the solution, but I'm not going to spend the next several hundred years writing a renderer in CUDA. Clearly the emphasis is for people like the authors of 3DSmax to start porting their code-base to make use of GPU computing. And if that happens, then clearly the solution for me is to pop some GPUs into the PCIe slots on the machines I build.
Basically, I'll leave my options open, but for now, I'm assuming GPU computing is not happening anytime soon w.r.t. rendering.
Yes, this is another good point. Total cost of ownership will be something I have to factor in, and may well mean that a corei7 will actually be the cheapest overall way to do things.
Yes, I think you are right. I threw that idea out there before I really looked into what was available.
I was saying that the intel option out laid earlier was using a £90ish board, when an older £35 board (G31) would have bought it in at a similar price as the AMD build. I'll let other people argue the pro's and cons when it comes to AMD vs Intel in reference to rendering
Well that isn't what you said now is it lol. Older G35's are.. G35's. G31's are G31's.
true though they can be had very cheap but they always look so naff I could never trust them.
Iv done a few builds with P31s and 33s and never had any issues with them.
Ok - so an update on this. Currently I'm working on these assumptions:
1) Total cluster cost is going to be ~£2000.
1) This cluster will be on 24/7, electricity is not free, therefore fewer very fast systems is better.
2) Software will not be free, therefore fewer very fast systems has a cost advantage.
3) Noise levels are irrelevant.
4) Looks are irrelevant - but keep it clean/simple. Machines may be stacked.
5) Graphics card is currently irrelevant, but may not be in the future (i.e. GPU rendering).
6) Storage is not too important. Optimise for speed/value rather than size.
So... this lead me towards an i7-based cluster of 3 machines, and bear in mind that I'll be overclocking each i7 up to ~3.2 GHz. Further stipulations are:
- Good-quality, reliable components that are also good value.
- Spacious chassis with ample cooling, not too expensive.
- Good + cheap i7 heatsink required.
This is my first iteration:
This comes to ~£630 ex. VAT.
So for 3 machines that comes to: £1890, which leave ~£100 for a single cheap monitor. I guess I'll also need a KVM switch, keyboard and mouse, so we're looking at just over £2000 excluding VAT.
Comments please!
The thing I'm really not sure about is the motherboard. I want a good quality mobo, but most of the stuff on X58 motherboards is totally unnecessary for this application. And currently they don't seem to come much cheaper than the above. Also, if Linux is the operating system, I'll probably switch to a cheap Nvidia graphics card instead because I believe they have better linux drivers.
Last edited by Fraz; 15-02-2009 at 01:54 AM. Reason: spelling
sounds pretty good to me bud except for the but ugly casei would spend a bit more and go for a coolermaster of sorts.
Fraz (15-02-2009)
A Coolermaster Elite 335 costs less than that and looks much less hidious.
Fraz (15-02-2009)
I'm quite surprised you can get all that for the money to be honest
The good thing is, the majoryty of that spend can be re-used even if the machines recieve a major overhaul. Realisticly you could just replace the processor in a year to get a significant boost in performance. Or even keep those systems and add two new graphics cards per box if the rendering software allows.
Pretty impressive all in all.
I guess if you wer buying 100 machiens instead of 3, you could build sample systems to work out a performance vs. cost vs. power usage model and figure out what was the best over all, but with anything less than 10 systems, the benefits of one build over the other are probably too small to make a real world difference.
Fraz (15-02-2009)
There are currently 1 users browsing this thread. (0 members and 1 guests)
Posting Permissions
Reply With Quote
