hybrid render slave

[MaRo]

Hi guys,

I'm thinking of building a small but powerfull headless render slave to nicely complement my main laptop (see my signature for specs). The idea is to chunk in as much as GPU power, with a nice processor (i7) and a decent amount of RAM (8Gb).

My question is about the GPUs. I have been using alternatively nVidia & AMD (former ATI) in the past so I do not want to start a flameware here, but currently I prefer OpenCL, and I consider the best implementation yet is AMD's one. Plus, the Vec5 architecture seems better to me than nVidia's scalar approach for massive parallel computing. So I'll go for ATI.

All in all, my question: would it be better to throw in several low end GPUs rather than one or two high end monsters.

I have the feeling that several Radeon 5830 1Gb (1120 stream units each) will perform better with LuxRender in hybrid mode (or SLG) than one big 6970 for example (1536 units).
But then, the questions of space/PCI slots/MoBo/ACPower unit/Heating are intimately linked.

I'd like your advice in terms of price/power/electric consumption/heating ratio...

Shoot at will!

PS: I posted that one in the general discussion, as it is aimed at LuxRender as a main use, but I did not want to put it in the GPU acceleration section of the forum, as it has nothing to do with the development. I hope it is the right place.

[SATtva]

With ATI/AMD and its current state of drivers you cannot have a truly headless (aka displayless) machine: for some obscure reason they require an attached monitor to run OpenCL apps. And this makes me sad. AFAIK, there is no such requirement for Nvidia GPUs/drivers.

[MaRo]

With ATI/AMD and its current state of drivers you cannot have a truly headless (aka displayless) machine: for some obscure reason they require an attached monitor to run OpenCL apps. And this makes me sad. AFAIK, there is no such requirement for Nvidia GPUs/drivers.

Ah. Couldn't we trick the system by opening a remote desktop?

More seriously, this is a bit annoying. I'd like to avoid buying a windows licence, and would rather stick with Ubuntu.

[SATtva]

Ah. Couldn't we trick the system by opening a remote desktop?

Nope, it requires physical connection.

[Zalamander]

The monitor connection is a requirement on Windows because of some WDDM restrictions afaik. You can't runOpenCL from a (Windows) Remote Desktop window either. However you can put a vga dummy plug (you can build it yourself with some resistors and a DVI-VGA adapter) on the GPUs and it will work just fine, you don't need a real monitor. As for the Remote Desktop restriction, you can use VNC instead. The restrictions do not apply for linux afaik.

[Zalamander]

And another thing: Fitting 4 GPUs into a system will create a huge amount of heat, and some fans are not designed to be packed tightly together. Also: Right now ATI devices are limited to 1/2 RAM for OpenCL, but this may well change (NV devices don't have such a limit anymore). As for rendering performance: Since the NVIDIA driver update their cards got a huge performance boost. So when you look at benchmarks for comparison, make sure to look at newer ones such as this one:
http://www.anandtech.com/show/4135/nvid ... -market/15

[MaRo]

Fitting 4 GPUs into a system will create a huge amount of heat, and some fans are not designed to be packed tightly together.

You're right, I searched for MoBos which could handle that many GPUs, and there are not much of them out there. And they are quite expensive.

So I will start "small", I bought two Radeon HD5830 Xtreme 1Gb, which IMHO give a good ratio power/price.
I'm now in the process of selecting a good socket 1155 mobo (to fit a Core i7 2600), with 8Gb RAM probably.

So I'm especially looking at the PCIE slots specifications. Most of the time, it's like 16x when one GPU is connected, then 2*8x when you have two GPUs
Some exist however that can handle 2*16x. But far more expensive. Is it worth?

Btw I've given up the idea of making this small rig headless. I have an old flat 15" handy, that'll be enough to configure the machine and keep an eye on the activity, whilst enabling OpenCL.

[SATtva]

The restrictions do not apply for linux afaik.

The last time i checked -- they applied. VGA dummy is a good idea though if somebody really wants it.

So I'm especially looking at the PCIE slots specifications. Most of the time, it's like 16x when one GPU is connected, then 2*8x when you have two GPUs
Some exist however that can handle 2*16x. But far more expensive. Is it worth?

Nope. Dade provided some benchmarks in the GPU forum section -- the differences between 16x and 8x was negligible.

[MaRo]

Thanks SATtva. That's a good point, lowering the cost of the mobo will allow a bigger quantity of RAM or a quieter case/power unit.

I'm changing the topic to headless hybrid render slave, as headless seems difficult to achieve (unless you are an electronics geek to make a dummy vga plug)

[Abel]

headless seems difficult to achieve (unless you are an electronics geek to make a dummy vga plug)

You're putting the bar for being an electronics geek pretty low.

According to http://www.overclock.net/folding-home-g ... -plug.html, it doesn't involve any soldering. Also no disassembling of the plug is needed, just stick some resistors in the right holes and you're done.