Rendering hardware

[HalifaxJ]

Greetings,

 

A friend and myself are looking to start a small Architectural Visualization company as a part time venture. We both have a few years experience using 3dsmax and are both intern Architects. I will be responsible for getting the hardware.

 

Initially, I was considering getting 3 machines with Intel processors and creating a small network. Rendering would be done through Backburner.

 

Posting on a local hardware board, a member told me to look into Nvidia solutions. I had heard of CUDA before, and was curious so I went reading. However, the information is not very clear at all as to whether their systems are intended to improve real time on screen rendering or speed up rendering times.

 

For example, the Quatro line of GPU is compatible with CUDA, and there are plug-ins for the Mentalray renderer which is now owned by Nvidia. Do the Quatro GPUs speed up rendering times or simply make the real time, on screen experience faster?

 

In my experience, graphic cards don't mean much for rendering. It all comes down to PCU power. So my question is whether a Quatro, through CUDA, is actually using the GPU to render or is it still my PCU?

 

On the other hand, they have the Tesla computer, which they advertise as a "cluster" within a desktop - true GPU computing. Will this speed up rendering times? Every article I can find seems to talk about scientific calculations.

 

In the end, I want to put together a small system and my end goal is rendering power and speed. As for on screen real time graphics, my Nvidia 8800GS is doing fine.

 

To recap:

 

1. Does a Quatro system render faster than a traditional multi-cpu network?

2. Is a Tesla system capable of rendering as a "cluster"?

3. What should I be looking at purchasing?

 

I'm not quite sure what the Nvidia Workstations do in terms of rendering as all the literature is not geared towards rendering and seems to be describing real time on screen issues.

 

Thanks,

 

Julien

[IceAged]

The GPU solutions you have listed will not benefit you in any noticeable way, since Max draws purely on the CPU for rendering. The Tesla system is, as you suggested, used mostly for scientific apps - at least for the time being.

 

If you want pure rendering speed and high performance, go for the Intel I7 cores. Of course, a good graphics card is a must for the workstation viewport performance - a standard GeForce gaming card will more than suffice, and will probably work out cheaper.

[AJLynn]

Forget CUDA, Tesla and GPU computing in general. It's not ready for prime time. No renderer you would actually use runs it, and whatever the rumor mill has churned up abut mental ray, it's not in the new version of Max.

 

A Quadro card has no effect at all on your renderer. In this economy, and for somebody getting started, it would be a complete waste of money. Heck, you don't even need a Quadro to run CUDA.

 

What you should look at is systems with a Core i7 920 CPU, 6 or 12 GB of RAM and a Geforce card. (If you're choosing components, a Geforce 9800GT is as good for what you're doing as anything else.)

[HalifaxJ]

Mental Ray is in Max 2009...

 

Some of the new CUDA packages are designed for Max 2009, and support Mental Ray according to the Nvidia site. That's where I'm probably getting confused. All the literature seems to say in some way that CUDA will hand off computations to the Quadro FX line of GPUs. Maybe I'm reading it wrong.

 

If we don't go QUATRO FX with CUDA, here's what I was thinking.

 

Main computer i7 920, 64bit OS with 12GB RAM, Geforece 9800GT. Vendetta 2 CPU Heat Sink as I'll be slightly overclocing. WD 640GB HD.

 

Two Q9400 machines with 4GB RAM and V2 Heat sinks as they too will be OCed.

 

Set up the system as a small network and run Backburner.

 

J

[AJLynn]

What I meant was, despite the speculation there is no support in Max 2010 for using mental ray with CUDA. CUDA isn't just something you add to your system and automatically all software can run on it. It requires software to be rewritten, and despite its having been around for more than two years there is zero useful software for it. Meanwhile nVidia puts up these grand pronouncements and hints at how fast CUDA is, which would be great if it actually did anything.

 

Personally I think the whole thing is vaporware and if it ever does amount to anything more than a curiosity and a few video game features, it won't be for years. Do not buy hardware with the expectation that it will be used with CUDA.

 

Your workstation idea looks fine but the Q9400 machines would have to be 1/3 cheaper than the comparable i7 to be worthwhile.

[HalifaxJ]

"Your workstation idea looks fine but the Q9400 machines would have to be 1/3 cheaper than the comparable i7 to be worthwhile."

 

Not sure I follow...the system is the 3 computers in a network.

 

Thanks for the info,

 

J

[ihabkal]

What I meant was, despite the speculation there is no support in Max 2010 for using mental ray with CUDA. CUDA isn't just something you add to your system and automatically all software can run on it. It requires software to be rewritten, and despite its having been around for more than two years there is zero useful software for it. Meanwhile nVidia puts up these grand pronouncements and hints at how fast CUDA is, which would be great if it actually did anything.

 

Personally I think the whole thing is vaporware and if it ever does amount to anything more than a curiosity and a few video game features, it won't be for years. Do not buy hardware with the expectation that it will be used with CUDA.

 

Couldn't have said it better! I wanted to say something like this since when I first saw this thread 8 hrs ago, but I was afraidf no one would believe me. I think you are right on the money.

[AJLynn]

Ihab- Thanks for the agreement, I was sure somebody would get all uppity over that comment

 

Julien- What I meant was, since the Q9400 is last generation, while it is less expensive it is also less powerful and will take up the same amount of space, time to maintain, software licenses and electricity. So it's the same overhead for less power. It needs to compensate by being significantly less expensive than an i7 alternative in order to be a good value decision.

[HalifaxJ]

AJlynn,

 

Thanks. Hadn't thought of it that way. However, 3 i7 machines are beyond our budget. On the other hand, we could start with 2 i7 machines...and keep adding slaves as we go.

 

Is the i7 920 (2.66Ghz I think) much faster than a C2Q9400 rated at the same clock speed? I understand they've eliminated the Front Side Bus which is what provides the significant (?) speed increase?

 

This being said, I can get a Q9400 system running for 800$, whereas the i7 is a minimum of 1500$.

 

J

[AJLynn]

At that price difference the Q9400 might be cost effective. See this chart to get an idea of render times. (mental ray and Vray are a bit more efficient in multi-CPU rendering so the actual score would be biased more in favor of the i7, which uses Hyperthreading, but nobody does a controlled comparison on mental ray or Vray.)

 

BTW they didn't eliminate the FSB - that would be impossible without a total redesign of the PC concept - but they did move the memory controller to the chip, which is more efficient, and made a bunch of other improvements.

[Jeff Mottle]

As an aside to the discussion about GPU rendering, there is software that is primetime for GPU rendering. http://studiogpu.com/ It's not the same quality as v-ray or mental ray yet, but the software has been used in full time production for several years. The company that created it has been using it on their own projects and just launched it a few days ago as a commercial package. I have seen it work in person several times and it's very fast! From what I have seen on the market thus far, they are the only people who have a product that is this far along and production ready.

[BOXXLABS]

rumor is that these guys are making a big announcement at NAB this week

[AJLynn]

Until I can download the demo, it's vaporware. Sorry Adam but we've seen and been disappointed by too many vaporware rendering technologies (anybody remember the Maxwell RC dark ages?) and it's not something to influence a hardware decision.

[BOXXLABS]

oh absolutely AJ. I agree 100% that its too early to tell anything concrete.

 

However, the tech is certainly not vaporware. It's real enough... Public demonstrations are rumored to begin directly - but private demos have apparently been ongoing for the past several months.

 

 

I've been involved at some level with hardware-based image processing that spans a decade - Blue ICE for AE, ARTVPS (for both renderman and mental ray), GELATO on NVIDIA, etc.

 

And now the Caustic Graphics approach...

 

This time, though, I think the "kinks" may be worked out to the point that this may be a truly viable technology for cinematic rendering.

 

I know several of the early employees of the company. They definitely have the right people and expertise to back up a very serious run at this.

 

Here's a preview: http://www.pcper.com/article.php?aid=694

Edited April 20, 2009 by BOXXLABS

[Jeff Mottle]

I'll be posting an interview with one of the Caustic executives this week, but the technology is only being released to developers this week. There is no public product yet. Probaby won't be for a while yet and Brazil will likely be the app that does the integration first given they were purchased by Caustic. Just to be clear, the Caustic card is not a renderer. It's an accelerator for raytrace calculations. It will speed up CPU renderers (who integrate the card into their renderers) and then GPU rendering products later on. The card required that people like Autodesk, mental ray, chaos group etc integrate the hooks into the Caustic card.

[AJLynn]

Great. Another proprietary hardware solution requiring developers to make proprietary mods for their software.

 

Remember how mental ray was supposed to run on GPUs because it was bought by nVidia? What has that done for us lately? Sorry guys, but I've worked in software and I know vaporware when I see it and this is still vaporware - and not particularly good vaporware at that. At least nVidia vaporware runs vapor mental ray instead of vapor Brazil.

[BOXXLABS]

...Remember how mental ray was supposed to run on GPUs because it was bought by nVidia? What has that done for us lately?...

 

 

To be fair, neither nvidia nor mental images EVER said they expected to see a product based on mental ray-on-hardware implementation anytime soon.

 

This has been (until very recently) a myth perpetuated by well-meaning but nevertheless clueless people engaged in "flights of fancy" and conjecture.

 

You are right that the ISVs - the software manufacturers - are going to have to embrace (or at least support) this new technology/product approach. But my bet is that if they (Caustic) can enable the types of workflows and speed increases they say they can, most folks won't have too much problem installing the hardware, drivers and a plugin or two to make this work.

 

And I have to disagree with your notion of "vaporware" in this case AJ. Respectfully, If you read the link (I posted in above) all the way to the end, you will see a video demonstration of a working prototype. In my book, this clearly liberates a technology or product from the marginalizing stigma of being called "vaporware".

 

Are you saying that you consider something to be "vaporware" until you personally can evaluate it at your desk?

 

 

Cheers,

 

Adam

[AJLynn]

Are you saying that you consider something to be "vaporware" until you personally can evaluate it at your desk?

 

 

Cheers,

 

Adam

 

That is exactly what I am saying. Well, I'd settle for somebody else I know personally and have met in real life having an evaluation copy and telling me about it (by voice, not email).

 

Show me the product. Until it exists, it's vaporware. Prototypes do not demonstrate reality. It would be unfair to call CUDA vaporware and not also call this thing vaporware.