Cycles render farm brainstorming

I’m looking into building my own custom Windows render farm for Cycles rendering in Blender. While I have a pseudo-render farm that I built/coded for Anime Studio, most of the computers I use for rendering with that program don’t have a CUDA-compatible graphics card (and cannot have one for various reasons, like, because it’s a laptop). So I’m looking at acquiring several computers that I can outfit with a Cycles-friendly GPU.

I’d like to keep things as cost-effective as possible. My first thought was to pick up several computers that are last year’s model, especially if I can find a computer like that has a decent CUDA-compatible graphics card. I have no problem assembling computers from scratch, but it doesn’t seem like a cost-effective move compared to getting an already-assembled desktop with acceptable specs.

From what I can tell with Cycles rendering, what matters is not the computer’s CPU speed, the onboard number of CPUs, or how much motherboard RAM the computer has; what matters the most is the video card’s speed and how much RAM is on the video card. Is this a correct assumption to make?

As an aside, I’m not interested in using an online render farm service. Each shot in the project I’m working on is going to be gigantic. A lot of shots will end up being over 1 GB, with all of the image sequences (and accompanying alpha transparency image sequences) that I’ll be incorporating. I’m going to be rendering 1920x1080 png files stereoscopically at 30 FPS. A rendered 18 second clip creates about 3.35 GB of PNG files. That’s just too much data to be shuffling around on my internet connection. So I want it all in-house.

Anyway, I’m looking for opinions/advice – specifically, on approaches I might take, hardware that I should consider, and bargains that I should look into. I haven’t decided on a budget or on how many computers I want; I’m still in the very early stages of figuring out the best path to take.

Thanks!

Hi, the problem with GPU rendering is the limited video RAM.
You have at least go for 3,4 or 6 GB GPU and the most cost effective way to this is one system with several GPU.
One case, one power supply and a board with 3 or 4 (very expensive) PCIe slots.
Than you can, depends of your budget, put 1-4 GTX 780 6GB or GTX 760 4GB cards in this system.
You can also start with one card and extend the system later very easy.
Here is a spreadsheet with all cards:

/uploads/default/original/4X/1/3/8/1388c64bc2707267be2c0434c1a56037057fc4ab.pngd=1400534233

Here is the benchmark for 3 Titan and changed tile setting (3 Tiles).

http://www.blenderartists.org/forum/showthread.php?331149-The-new-Cyles-GPU-2-70-Benchmark&p=2656914&viewfull=1#post2656914

The benchmark file is on the first page, try it on your system.
One important advantage of one system is your electric bill does not explode as with 4 or more single systems.
For example, one big power supply is much more effective than 4 small cheaper power supplies.

Cheers, mib.

If you want a GPU based rendering farm, than you’ll want something like this perhaps:

http://www.pccasegear.com/sc/gyT - Prices are in $AUD, which will be more than most other places :stuck_out_tongue:

The important part to note is, you really do want to go with 6GB GPU’s, and multiple GPU’s per PC. In this case $1400 is spend on 2 6GB GTX 780’s, and the rest is on an 8GB RAM, Dual Module AMD CPU and a good PSU. You don’t need a powerful CPU for GPU rendering.

Thanks, mib2berlin. Currently, I have a EVGA GTX 580 (3 GB) on Windows 7 Home. I ran the test and got 1:56 using Blender 2.69. The Blender 2.70 alpha as well as a Blender 2.70 “Fastest Build” versions both took a lot longer – it was around 2:10 IIRC.

I have one other video card slot free on my motherboard. I’ve got a beefy power supply and could handle another big video card. The Titan seems to be overpriced for what it is. (Two 3GB GTX 580s render faster – and are cheaper – than one Titan.) Will this particular ASUS 6GB GTX 780 play nice with my EVGA GTX 580? http://www.amazon.com/Thermal-Direct-GPU-Contact-Graphics-GTX780-DC2OC-3GD5/dp/B00DWV3NM6

Hi zeealpal – the link you posted doesn’t really go anywhere on that site. Could you describe what hardware that link was supposed to show?

Hmm, the link was meant to show an example build for a GPU render farm, screenshot below


Ah, OK – it does indeed take me to that screen. I wasn’t sure exactly what I was looking at. Thanks!

Hi, the GTX 780 6 GB should render the scene in 1.20 ~ in 2.70a but iirc newer cards are better on heavy scenes.
VRAM is not added with more GPU, on very heavy scenes you can use only the 6 GB card but if the scene fits in to 3 GB you can render with both cards.
To get the most out of you 3 GB card, drive the display with the 6 GB card (System use 3-400 MB).
I see no reason why the two EVGA cards should not work very well together, and fast.

Next step is, change your GTX 580 with upcoming GTX 880 8 GB. :yes:

Cheers, mib

Thanks for the info. I’m actually having trouble finding a place that sells that GTX 780 6 GB model for a decent price (in US dollars). Amazon only has it for $849, there’s none for sale on Newegg or on eBay…even EVGA’s web site isn’t selling it (there’s a button to be notified when it’s on sale): http://www.evga.com/Products/Product.aspx?pn=06G-P4-3785-KR There’s no problem finding the 3 GB model, though.

I see a Zotac 6 GB GTX 780: http://www.amazon.com/zt-70210-10p-gtx780-384bit-pci-express-displayport/dp/B00K9Q6C48 But that’s not EVGA – I’m not sure if there would be problems if there were two video cards from different manufacturers in the same system.

Actually, I may pick up another GTX 580 with 3 GB. It seems to be a much more cost-effective option than the GTX 780 (3 GB or 6 GB model), and my rendering speed will double. I can always pick up a 6 GB card for another computer later on, once everyone’s using the 8 GB model…