[piroshki]

Hi All, and NextLimit

I'm assuming everyone saw the announcements by Apple a few weeks back concerning their OpenCL project? From what I know it looks like the compute power on graphics card is about to go through the roof, with many of them having much more number-crunching ability that our actual CPUs... So this begs 2 questions:

- Are you guys looking into OpenCL, and how difficult would it be to make sure that Maxwell (and I'm sure RealFlow) leverages this stuff as soon as it becomes available?

- What does a render engine become when the computing power increases so much that render times are insignificant?

Cheers to all,

T.

http://news.cnet.com/8301-13579_3-9970617-37.html
http://en.wikipedia.org/wiki/OpenCL

[piroshki]

I see I'm not getting any responses...

So what if you had not 2 cores, not 4, not 8, not 16 (top benchwell entry), but 256!!!

Well, its here. And its only going to get better.

Some will nit-pick that these aren't true CPU cores - but imagine being able to dump a whole bunch of parallel computing onto this large number of processing cores...

http://www.anandtech.com/video/showdoc.aspx?i=3334

[messire]

i'm with you piroshki cant wait to see !!
N

[Bubbaloo]

I would like to hear what the NL guys think, too.

[Carl007]

bumping this thread with a link

http://en.wikipedia.org/wiki/OpenCL

.... lets any application tap into the vast gigaflops of GPU computing power previously available only to graphics applications. OpenCL is based on the C programming language....

Sounds great! Will this be beneficial for nextlimit applications?

[Mr. Gog]

To have a complete idea of Open CL, take a look of Grand Central:

http://blogs.zdnet.com/storage/?p=334

http://bits.blogs.nytimes.com/2008/06/1 ... side-down/

http://www.apple.com/macosx/snowleopard/

[Carl007]

Thanks Gog.

[yanada]

I would like to hear what the NL guys think, too.

INDEED

[Cadhorn]

I think it's fair to say that *every* maxwell user is wondering if this cuda/opencl stuff can be used to speed up rendering.

I'd love to hear from NL about this. I'm sure there's more to the story.

Maxwell Render would be the undisputed #1 render engine if the speed could be kicked up 100x (or even 10x).

[sam7]

Finally, that would be a reason for me to upgrade to Maxwell Render 2.0 if it at least parts of the rendering pipeline would make use of OpenCL.

[msantana]

Just to add to the Open CL speculation....

This is by no means scientific or double-checked, but I think it is at least a starting point to get some idea if using GPUs might be a good idea.

Suppose that Maxwell scales linearly with gigaflop performance (roughly billions of floating point instructions calculated per second) which I think is reasonable since we have seen it benefits almost linearly from multi-threading.

Suppose you have a 5300-series "Clovertown" 8 core 3Ghz Machine. It is roughly benchmarked at 76 gigaflops (38 each processor) according to this wikipedia article:

http://en.wikipedia.org/wiki/Xeon#5300- ... vertown.22

If you add to that machine a QuadroFX 4500 card, it is roughly rated at 200 gigaflops, according to Nvidia's own marketing literature, which can be found here:

http://www.cad2.com/pdf/tesla_technical_brief.pdf

So, assuming Maxwell is coded in to take advantage of the card, and you believe the above literature, it could mean the machine would be capable of 276 gigaflops, a 163% increase in performance on your renders. An hour render would only take 17 minutes...

Even if Next Limit only achieved a 30% increase in performance I would be sold in using a GPU card for this.

My 2 cents of pure speculation.

[dmeyer]

how about multiple GPUs?

[msantana]

how about multiple GPUs?

Well, if you are really interested, Nvidia has been selling these for a while:

http://www.nvidia.com/object/tesla_s1070.html

The equipment above is a GPU subsystem that you connect to a Windows or Linux machine and can perform computations on.

And the ones in the link below are not as expensive, they are just systems with two QuadroFx cards (two 5600s) that are configured in parallel, i.e. they have been optimized to share information and the memory throughput of the communication between the cards has been enhanced (But you still need a Windows machine to connect them to, i.e., they are not standalone renderboxes). This system would definitely be faster than just plugging in two QuadroFx cards into your own PC:

http://www.nvidia.com/object/tesla_d870.html

But then, I think it becomes a question of what is less expensive, an x86 renderbox or the above. (For that amount of money you can get probably a 16-core system (4 processors with 4 cores configured in two computers) with at least 16 gigs of RAM, i.e. 8 gigs per system) And this will depend only on how Maxwell utilizes the GPUs.

Let's see if there is any interest by NextLimit to pursue this.

[dmeyer]

But then, I think it becomes a question of what is less expensive, an x86 renderbox or the above. (For that amount of money you can get probably a 16-core system (4 processors with 4 cores configured in two computers) with at least 16 gigs of RAM, i.e. 8 gigs per system) And this will depend only on how Maxwell utilizes the GPUs.

Let's see if there is any interest by NextLimit to pursue this.

The issue with a lot of these solutions is that they require custom code to leverage the GPU power. If NL were to write a plugin that really takes advantage of Quadro's to speed rendering...even more than a bunch of general purpose CPUs, I'd be all over it.

[msantana]

The issue with a lot of these solutions is that they require custom code to leverage the GPU power. If NL were to write a plugin that really takes advantage of Quadro's to speed rendering...even more than a bunch of general purpose CPUs, I'd be all over it.

True, the custom solutions from nVidia (the Tesla products) require programs to be coded with cuda. So this implicitly assumes that Maxwell would be rewritten in cuda. But hey, if we already brought up grand central and OpenCL we might as well talk about cuda.

Also, since you brought this up, it is worth mentioning that I think that is one of the reasons Apple didn't adopt cuda and is proposing OpenCL. They want this to be a standard that other companies adopt as well, and it is not only Nvidia's (plus to a point they would be competing with nVidia for hardware sales if they supported cuda directly).

cuda works on Windows and Linux, but as you mention, requires a complete rewrite of the program. I think it has had little adoption, at least in the graphics industry, because of this. People don't want to tie themselves to a specific hardware platform, even if it s open.

Next Limit has been rewriting Maxwell for Mac OS X (porting it from C++ to Objective-C). And that I think is more promising than porting it to an even smaller installed base with cuda. But at least they have some experience in porting the code from C++ to another language.

In short, I don't think there is any way of knowing what's better until implementation is done. Let's see if Next Limit considers implementing OpenCL or other GPU-based rendering technologies. As I said, I think once Maxwell is ported (if it is ported at all) it will be a question of what platform is cheaper to render on.