need suggestion on buying a motherboard supporting SLI (1150)?

[Master]

Hello everyone.
I'm currently in a situation where I can not upgrade to the 1151 or newer architectures.
I want to buy a motherboard now, that for the next 2 years at least fulfills my needs.
My prime need is being able to do SLI , two way or 3 three way. the intention is not for gaming but Deep Learning related tests. (Actually I want to be able to install as many graphics card as possible to benefit from it in my computations and refrain from buying a new expensive one !)
So why do you suggest buying, all things said ?
I myself like Asus more, but Gigabytes and MSI are fine as well (these are the brands that are more trendy and available in my country compared to others)
Thank you in advance

 

[Ja.KooLit]

the only lga 1150 board that can have alot of pci slot are those with plex chip like asus maximus vi extreme or gigabyte m1 sniper.

i believe msi and asrock offer same with plex chips but i dont remember their mobo names.

 

[Master]

the only lga 1150 board that can have alot of pci slot are those with plex chip like asus maximus vi extreme or gigabyte m1 sniper.

i believe msi and asrock offer same with plex chips but i dont remember their mobo names.

Thank you very much
by the way are all of these PCIexpress slots the same? I mean are they all 16x ?

 

[Ja.KooLit]

Thank you very much
by the way are all of these PCIexpress slots the same? I mean are they all 16x ?

actually no.

like for example on Asus Maximus VI extreme, configurations are:
x8, x16, x8, x8
https://www.asus.com/Motherboards/MAXIMUS_VI_EXTREME/

gigabyte z87 sniper 5
http://www.gigabyte.com/products/product-page.aspx?pid=4487#sp

if you are asking about slot if it is x16, then yes all are x16 but they will not work at full x16. only x8.

By the way, asus also has WS boards but all have PLEX chips for quad PCI-E slots

 

[Jetster]

I will be selling my Gigabyte Z97X Gaming G1, wait your in Iran? Never mind

 

[jboydgolfer]

I was able to do GTX 970 reference design SLI on my Asrock Extreme 4 Z97, and if I recall correctly there were still some X4 & x16 slots left over. It was under $100

 

[newtekie1]

Do the cards have to actually be in SLI, or can they work independently?

All the 1150 motherboards that I know of only allow 2 way SLI maximum.

If the cards can work independently, and don't actually have to be in SLI, then just find any motherboard with a lot of PCI-E x16 slots and cram as many cards as you can in there. I believe AsRock has some really inexpensive boards that were targeted at Bitcoin mining that had a lot of PCI-E slots but not a lot of other features to drive up the price.

 

[jboydgolfer]

The Asrock Extreme 4 I posted above officially supports quad SLI and three-way crossfire. But what it officially supports and what you can actually do are two different things entirely. My guess is those claims would only be attainable on a test bench or in open air case

On an entirely separate subject. I like the board a lot, it has great UEFI and tons of options ,great overclocking too




And a side note.....dont hold it against the MoBo, but it is "Toms HW approved "
so its approved by a terrible forum site, with more Ad's than content.

 

[newtekie1]

The Asrock Extreme 4 I posted above officially supports quad SLI and three-way crossfire.

The support of Quad-SLI is a little misleading. It only supports 2 cards in SLI. But if you use dual-gpu cards, then you can do quad-SLI on it.

I have an Extreme4 and an Extreme6, both are limited to 2 cards in SLI.

 

[okidna]

@newtekie1 is right, you don't need a SLI supported board if your main purpose is to build a GPU-based deep learning data trainer.
I myself use my old 990FX mobo with 2 x GTX 980Ti in my lab to train mainly CNN and my friend sometimes use it to train her RNN without using SLI bridge (I mostly uses Caffe and Theano and recently Keras and Tensorflow).
From my experience, having the PCI-E to run at maximum bandwidth will help if you're transferring a lot of data from host to device(s), for example when you preprocessing or training a lot of images (when I say "a lot" it means at least like ImageNet dataset).

As for multi-GPU training support, there's no NVIDIA official support for running the training in multi-GPU. You have to tweak your code and your NN to be trained with more than 1 GPU (Theano can do this easily by sending the command with "device" attribute), or if you want it to be more simple you can try to use DIGITS to train multiple networks on the same data set in parallel, or train the same network on multiple datasets in parallel.

 

[Master]

Thank you very much guys, I really appreciate your helps , they were really informative