how to build a cuda render server?

[1832vin]

say if i wanted to build a CUDA render server (with 30 GPUs) that takes requests over (e.g cat5e) and then renders my 3D animation

how would i go around building something like that?

 

old:

{

   super computers are the pinnicle of computing, and why stop at desktop and servers?

   if you know about supercomputers (ie, crosslinks names, how they work, distribution,diagnosis, fixing, analysis, and OS (ie supermuc or redhat)  where did you learnt about them?

 

   because in manyway a paralleled compute device is much more different compared to desktops, since most of the time, desktop CPU are the one that does the most job, but 

   the case of a blade housing, there are multiple types of nodes, and different setups are better for certain tasks(ie cloud storage solutions, simulations, lookups)

   and i would like to learn all of those

 

   but the problem is, there isn't much

   the pages i found on a google search was mostly "super computers are superfast!" and "PETAFLOP" and not much useful things

 

   so if you know anything, where did you learn it?

}

[SIGSEGV]

Well there are clusters, which are just a load of computers working as if they were one

[SuperPug]

''super computers are superfast!'' <- Thats pretty much as far as my knowledge of supercomputers goes.

[Jade]

OS is typically custom or redhat, but it really depends on the one you're referring to.

They work in a cluster; a group of computers are working on the same dataset to compute the outcome.

Diagnosis: The 'overseeing' machine could see a specific device fall off the cluster and identify which, diagnosing and fixing would be similar to a desktop PC's after locating the failed device (again, via the 'overseeing' software.)

[APersonWithABrokenChair]

Here:

https://www.youtube.com/watch?v=unDqlQA9rU4

[1832vin]

Here:

no, that's not enough, say i would like to build a render server, consisting 30 GPU, how should i do it?

 

that video was more history, than what is

[dalekphalm]

no, that's not enough, say i would like to build a render server, consisting 30 GPU, how should i do it?

 

that video was more history, than what is

Check this out: Should be a good place to start:

http://arstechnica.com/information-technology/2014/05/how-to-network-lots-of-dumb-computing-muscle-in-a-fast-efficient-render-farm/

 

EDIT: Nice Thread and title edit... well at least it makes your post easier to understand and much more reasonable

[Jarsky]

I believe this is what you're looking for:http://devblogs.nvidia.com/parallelforall/how-build-gpu-accelerated-research-cluster/

 

essentially you would build boxes stacked with nvidia gpu's and have 1 spare pci-e slot on each computer for an InfiniBand card for them to communicate.

There are some rackmount cases which take 2 motherboards so essentially you could have 2 machines in each rackmount.

 

I think the hardest thing, would be getting software to leverage all the machines as a single cluster.

[1832vin]

I believe this is what you're looking for:http://devblogs.nvidia.com/parallelforall/how-build-gpu-accelerated-research-cluster/

 

essentially you would build boxes stacked with nvidia gpu's and have 1 spare pci-e slot on each computer for an InfiniBand card for them to communicate.

There are some rackmount cases which take 2 motherboards so essentially you could have 2 machines in each rackmount.

 

I think the hardest thing, would be getting software to leverage all the machines as a single cluster.

i've already got the software, it's the hardware that i can choose from having 3 very powerful cards(titan black), or 25 weak cards (750ti), oh, btw, if i go with the 25X750ti, i can render stuff 8X quicker, but with less memory

[Jarsky]

i've already got the software, it's the hardware that i can choose from having 3 very powerful cards(titan black), or 25 weak cards (750ti), oh, btw, if i go with the 25X750ti, i can render stuff 8X quicker, but with less memory

 

OK thats much easier. theres a big difference between the 2.

 

3 powerful cards, you can just use normal consumer software.

If you go 25-30 cards, you're going to have to cluster machines. the most effective way of clustering is via InfiniBand communication (that bandwidth)

 

From a pure cost, power and ease of setup, 3 titans is going to win hands down against 25 x 750ti's.

 

As far as speed, 3 x Titan Blacks would have over 8,000 CUDA cores, where as 25 x 750Ti's would have 16,000 CUDA cores, so effectively it will only be *twice* the speed.

If you work with that much video you would probably be better just setting up 2 seperate rendering machines, that you can render different videos on in parallel.