850 Watts enough for 2 x GTX 970 ?

[reckless]

What do you guys think of the Gainward GTX 970. Its 15% cheaper than the Zotac reference card.
I have never heard of this company before. But if there is nothing wrong with Gainward,
then I dont mind getting it. Use the extra cash on other parts.

Just posted the Gainward thread here - http://render.otoy.com/forum/viewtopic.php?f=9&t=45683

[reckless]

This is specifically about Gigabyte GA-Z97X-SOC Force + i5 4690.

Just learned something new about PCI lanes between the CPU and PCI slots. i5 4690 support only 16 lanes and this Mobo can support up to 4 GPUs.
Which in my mind means that 16x will come down to 4x in a 4-GPU config.

If my assumption is right then would this affect Octane's performance in the future when I add more GPUs ?
Or does this apply to only gaming users who require a much quicker data transfer coz of FPS and all ?

Initially I'll add only two 970s.

Learning something everyday coz of this build
hahaha.. have finally come down to what Smich had suggested

[glimpse]

yes it's good to read some reviews
(TomsHardware/Anandtech are my favorite places for that =)

What Gigabyte done with this gen OC board (Z97) to lower the price - they don't include PLX chip - & thus You get lower lane count to PCIe slots =)

slightly older, earlier model : GA-Z87X-OC Force has PLX & it supports x8 x8 x8 x8 mode =)

[reckless]

So you mean this lane distribution does have a negative impact on Octane render performance ?
1x or 4x is slower than 8x or 16x in terms of GPU rendering ?

[glimpse]

So you mean this lane distribution does have a negative impact on Octane render performance ?
1x or 4x is slower than 8x or 16x in terms of GPU rendering ?

Rendering itself is not effected by upload speed..-once scene data is on VRAM speed is going to be the same whether that data came from x1 or x16..

if You're trafficking small amount of data in & out - You should notice little to no difference & only if You're taxing GPU to it's max..then it's another story..

haven't test this out, but I believe of the core texture functionality should be impacted by lower speeds =) mabe someOne made any tests?

p.s. something to keep in mind (recently read on the article about storage, but it's mroe or less in the theme here =)

"..with the help of PCIe switches it's possible to grant all devices the lanes they require (although the maximum bandwidth isn't increased, but switches allow full x16 bandwidth to the GPUs when they need it). "

so in simple words..when You have a switch it gives a possibility, but not the bandwidth.. - so if you're limited by CPU, You are..switches only hook all the devices..so in case when You need to drop all the information to one GPU it would get full lane count instead of 1/4th..

so as I look to that..as I understand..if You have the same CPU on board with PLX & without.. there would be no difference if You would use 4x cards =) but the difference would come if You would use 1 card, (but the rest would be idling) - got the point? it's switch, like a 'crosroad'.. not some sort of a 'factory' that makes more "bandwidth" =)

[reckless]

ok so it doesn't affect the rendering speed much. But data transfer to the GPU gets slower if you have lesser lanes. Got it.

Just checked wikipedia to get some PCIe speed numbers. And going by this screenshot it seems a 12GB GPU will get filled
with data in about 4-5 seconds (each slot using 4 lanes x 985MB/s x 12GB). Please correct me if I am wrong.

Capture.PNG (10.03 KiB) Viewed 3198 times

Got your PLX explanation. It adaptively allows cards to use 16x lanes, when they need it. But in a 4 GPU config that doesn't really work, like you said.

[sikotik13]

From what I gathered regarding PLX in general, it actually the opposite, the entire point is speed increase over multi-GPU. In systems without PLX, the lanes are limited in such a fashion that after 2 cards (usually at most) PCIe starts droppiing, so usually it'll turn out something like 2 ports at x16 with 1 port at x8, or 4 ports at x4. The PLX allows for you to have x16 on all four lanes, it just isn't increasing the bandwidth past the current speed (i.e it's not hitting PCIEx16 gen 4 or something). This is at least how they tend to present it, and information regarding it is often somewhat cryptic. No idea how much this affects speeds really, as far as Octane is concerned, but researching this was the basis of the parts I selected for my build.

[acc24ex]

I'm running two 590gtx + 580gtx on corsair 950W (that's 5 gpu cores), along with overclocked from i7 950 from 3ghz > 4ghz - and surprisingly stable .. win 7 clocked it at 7.8 (out of 7.9 top)

- more VRAM is always recommended, and it is a shitload of frustration when you can't squeeze in those additional things near the end of a project, but you can do it with a lot of optimization from the start, instance a lot, and try and optimise geometry etc.. I built a few parks with 1.5 gb limits, resolutions up to 1920x1200.. and I'm pretty satisfied with those so far, prints come out just fine..

- although it seems octane will do cpu+gpu rendering with 3.0 (something that they said they're never gonna do..) - so..

[smicha]

I am not an expert of PLX but correct me if I am wrong:
i5 4690 supports up to 16 PCI lanes so in 4 way gpu setup we'll always have x4, won't we?

"Max # of PCI Express Lanes
A PCI Express (PCIe) lane consists of two differential signaling pairs, one for receiving data, one for transmitting data, and is the basic unit of the PCIe bus. # of PCI Express Lanes is the total number supported by the processor."

After conversations with Piotrek (the guys that soon will benefit from watercooled 2xZs) I got to know that for heavy animation scenes, that require reuploading geometry for every single frame # of PCI lanes and x16 vs x4 matters. So if you are using (or going to use) Octane for animations maybe it's better to look for full 2011-v3 architecture (where my winner is X99-UD4P):

'GIGABYTE X99 motherboards feature a unique PCI Express design that utilizes 100% of all 40 lanes from the CPU when in a 4-Way or 3-Way graphics configuration. Standard designs limit all 4 main PCIe lanes to x8 (64Gb/s) bandwidth, but thanks to the onboard external clock generator combined with the direct connection of one of the x16 lanes to the CPU (switchless design), GIGABYTE X99 motherboards can unlock the full bandwidth available and provide the best graphics bandwidth possible to the user.'

But to get 40 pci lanes you'd need 5960X or 5930K

[smicha]

And here is something I don't get (asus x99 ws)

40-Lane CPU-
7 x PCIe 3.0/2.0 x16 (single x16 or dual x16/x16 or triple x16/x16/x16 or quad x16/x16/x16/x16 or seven x16/x8/x8/x8/x8/x8/x8)

28-Lane CPU-
7 x PCIe 3.0/2.0 x16 (single x16 or dual x16/x16 or triple x16/x16/x16 or quad x16/x16/x16/x16 or seven x16/x8/x8/x8/x8/x8/x8)


28 lane cpu and x16/x16/x16/x16?