Xeon PHI & Vray?

[Dariusz Makowski (Dadal)]

Heya

Will it work?

I fancy 2 of those in my rig...

http://www.tomshardware.co.uk/xeon-p...iew-32565.html

[vlado]

Right away it won't work. We would love to take a look at this, but we don't have the card here and I don't know how to get it.

Best regards,
Vlado

[TheLooopCGI]

I think Xeon Phi is an absolute must-have (provided you're okay with only 8GB of memory of course) since it's x86 and the power output is at 1TFLOPS. For comparison a Core i7 CPU is like 50GFLOPS "only". One such card is like a whole render farm...

[Dariusz Makowski (Dadal)]

I hope that u will be able to swap data and use PC ram... or at least have 2x PHI and they could share their memory so u end up with 16 gb of ram... For me a minimum is 24 gb atm

[Morbid Angel]

give it a few years guys...

[theedge]

I see That the Phi runs on Linux. Shouldn't this make it easy to run vray as there is already a linux version of it?

[vlado]

I see That the Phi runs on Linux. Shouldn't this make it easy to run vray as there is already a linux version of it?

Hopefully, yes. I don't know enough about it yet though.

Best regards,
Vlado

[3LP]

Yeah looks great, but still, if your saying that 1 i7 is doing "only" 50GFLOPS, isn't 1TFLOPS only twise the speed? So basically 2 computers, that's not really a whole render farm...
I hope I'm wrong

And what about the new cuda 5?
https://developer.nvidia.com/cuda-toolkit

I saw that :
"Direct communication between GPUs and other PCIe devices"

Does that mean that cards will be able to interact with each other and that ram could be shared between cards?
Like multiple cpu on the same motherboard share the same ram?

Thanks

Stan

[vlado]

Yeah looks great, but still, if your saying that 1 i7 is doing "only" 50GFLOPS, isn't 1TFLOPS only twise the speed? So basically 2 computers, that's not really a whole render farm...
I hope I'm wrong

The difference between 50 GFlops and 1 TFlop is 20 times, not 2 times

And what about the new cuda 5? https://developer.nvidia.com/cuda-toolkit I saw that:
"Direct communication between GPUs and other PCIe devices"

Does that mean that cards will be able to interact with each other and that ram could be shared between cards? Like multiple cpu on the same motherboard share the same ram?

Unfortunately, no.

Best regards,
Vlado

[Dariusz Makowski (Dadal)]

What u mean Vlado 'no'. They say that GPU's will be able to communicate with each other and access memory with no need of CPU interaction and so on.
http://www.youtube.com/watch?v=MVMjO...ure=plpp_video

Thanks, bye.

[vlado]

What u mean Vlado 'no'. They say that GPU's will be able to communicate with each other and access memory with no need of CPU interaction and so on.
http://www.youtube.com/watch?v=MVMjO...ure=plpp_video

One, this is not an automatic process in the same way that multiple CPUs can access the same system RAM (so a programmer must do tedious bookkeeping to remember which parts of a data structure are in the memory of which device; in our case, this would mean rewriting V-Ray RT GPU from scratch and throw away all the work we have done so far). Two, it is a lot slower than accessing the RAM of the same GPU (because the PCI bus is still much slower than on-board memory).

Best regards,
Vlado

[Morbid Angel]

http://en.wikipedia.org/wiki/Giga-

[DEPTH]

Seems like we can put 4 xeon phi card in the same workstation
- http://www.novatte.com/products/xeon-phi/m40phi
Maybe the distributed rendering performance may be A LOT better on PCIe than on the Gygabite network.

[Dariusz Makowski (Dadal)]

Ah right so thats interesting... I hope you guys would find some kind of solution for that :s

as to PHI workstation do you think that memory ram could be sharable or expandable ? Its not GPU so I guess it will use Vray production render so it should be also able to access motherboard ram ? Not sure how much speed do we get from pci3 tbh I have not checked it for a while now...

As to testing PHI u can rent server from them maybe? :] http://www.novatte.com/products/xeon...i-coprocessors

Thanks, bye.

[vlado]

So we had a brief look at the Xeon Phi documentation; some conclusions that we can make (although of course, we need to test them with the actual hardware):

*) For the moment, only Linux hosts are supported. So forget about 3ds Max right now. Support for V-Ray for Maya, V-Ray Standalone and V-Ray RT CPU seems possible and fairly easy, using the existing framework for distributed rendering (so the card would appear as an extra host available for DR).
*) It should be possible to port the full production V-Ray; no limitations like with GPU rendering.
*) Whether the card can access directly the main RAM, I don't know - but it can access directly files from the main machine, like textures, proxies etc.

In any case, it will take some time to research all this, so don't expect any results tomorrow But we will certainly look into it.

Best regards,
Vlado