Hardware Acceleration

[TimothyDel]

As one who is getting older and not wanting to wait for even the most beautiful V-Ray renderings, I was wondering about internal or external graphic processors designed to crank out renderings in a fraction of the time we’re used to. I understand a rack of Blade servers would create a nice renderfarm but the cost is prohibitive.

I was looking at Art VPS’ RenderDrive. Not a cheap solution and proprietary to their materials. However, their Pure card was a few thousand dollars with very good results. Physics cards also exist for only a few hundred dollars (and up) for gamers to simulate real-world object interactions. Does any “render card” or GPU exist to finally get large renderings accomplished QUICKLY?

There is one card that does exist for this purpose but there is no software written for it yet. The card’s approximately $1,300.00. It would need someone to “simply” write code in C the instructions to tell V-Ray to off-load from the CPU(s) to these GPUs. It would then work similar to the Pure card. Then I was thinking if this was accomplished how difficult would it be to stack them in an external case and work in parallel on one frame? If the client will pay just add more cards.

This is a large global company. However, the card’s technicians have not allocated time to do this and are willing to “farm it out”. A large CAD company is also looking to incorporate this technology.

Any ideas?

Sincerely,
Timothy
deldesign@aol.com

[RErender]

Maybe, someday...

http://www.chaosgroup.com/forums/vbu...ghlight=gpu%2A

[vlado]

We have looked into hardware/GPU raytracing in depth for the last year; our final verdict is that current technology is still too limited in this regard. All of the solutions available at this time focus on raw calculation power and on the SIMD data model (single instruction/multiple data). This works well for some types of calculations (e.g. large matrix operations, or physics simulations), but this model fits poorly with raytracing, which is better suited for the MIMD model (multiple instructions/multiple data), which you get with multi-core CPU systems.

Another severe problem is the passing of data back and forth between the CPU and the GPU/specialized processor, which is way too slow. To be effective, such a solution would need to run entirely on the GPU or the additional processor, however the developer is at the same time serverely limited as to what can be done on the GPU. Many solutions claim that you can "just code in C", but this is not so easy and there are many restrictions to what you can actually code.

For example, a highly optimized simple raytracer running entirely on a high-end GPU (to avoid passing of any data), is about twice as fast as a single CPU core - and at the same time is severely limited in functionality. Add another CPU core, and the speed is the same. Take a four-core machine and the CPU solution is already twice as fast as the GPU, without any functional limits.

There are some interesting developments towards real MIMD solutions though and we are eagerly waiting for them

Best regards,
Vlado

[Paul Oblomov]

But what about limited acceleration for example for glossy effects ? Kinda multipass when CPU render everything, except GPUs glossy.
If there's any test you, vlado, done maybe you can show some of 'em ?

[vlado]

The problem with a 'limited' solution is that there needs to be a constant flow of data from the GPU back to the CPU with the calculated results. It turns out that this is very slow - enough to make the whole thing pointless.

Best regards,
Vlado

[Paul Oblomov]

So, even with PPT we'll have no acceleration ?

[vlado]

<shrug> With the current state of technology, the acceleration would be roughly comparable to that of one additional core in the best case. So lots of effort for not too much...

Best regards,
Vlado

[Morbid Angel]

That suks...
but then again, i haven't seen a studio yet where most of the hardware was even up to date. Most of our equimpent is 2-3 years old. It hurts to work on it...
Getting some gpu acceleration is soo far out there...

[^Lele^]

Ok for GPUs, Vlado, but what about these things?

http://www.clearspeed.com/products/overview/

They already work with Mathematica and mathlab, and seem to be extremely apt at math calculations, without the GPU limitations.
The gains there seem Humongous.

[vlado]

There's the problem - raytracing is not just a lot of math calculations. It is probably easy to imagine that a computer program is just calculations, but this is not true - there are also a lot of logical operations, program flow control and data access. Even GPUs are extremely fast for pure calculations (matrix mutliplications and such); problem is, raytracing is also a lot of logic (conditional branches) and a lot of random memory access (you need to go through all those million triangles in the scene somehow); while processors can calculate at blazing speed, RAM access is still nowhere as fast - and there is no way around this.

Further on, all these boards currently rely on the same small chunks of mathematical operations being performed over large amounts of data, and this is where their speed comes from. Not so for raytracing - nearly every ray requires its own unique set of calculations.

As far as raytracing goes, something like the Cell processor makes a lot more sense, where you have individual cores capable of following their own programs independently from the rest (although currently the problem there seems to be with the slow data access), however we expect to see lots of exciting development in that direction.

Best regards,
Vlado

[RePlay]

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

It looks like Nvidia's Tesla card is the beginning of this process but Vlado is right, by the time you buy these specific cards for applications with bottlenecks in other components, you are better off buying the extra cpu power. Also, remember the power of distributed rendering of Vray... it is and has been a lifesaver for me in the past.

A Quad core these days runs very cheap. Cheap enough you can buy 5-7 Quad cores for the price of one Tesla Card. This would supposedly upgrade your DR farm considerably.

Instead of focusing on technology that is not widely integrated, I would rather Vlado works on physical materials and such things as to streamline the process of setting up the perfect render. Vray Realtime would be very handy and I'm sure it's around the corner.

Anyway this is just my opinion and thoughts

[dimo]

Hehe,

I have messed around a little with CUDA for the last couple of weeks and I found it really hard to work with. No objectoriented programming and even no recursive function calls. For a programmer like me who tends do it quick and dirty the code ends in a mess . But for particle calculations it is a good thing.
I suppose it would be also very good at accelerating AURA or FUMEFX

Best regards,

Dieter

[joconnell]

The only bad thing about these cards is that they're going to be very specific to applications that support them so you might spend the cost of a render node on one of them and only get it's benefit in a single package - if you just buy another render node you can use the cpus to render anything.

[mike_kennedy]

When I worked with a dealer years ago we took a close look at the Art solution and while is was fast it was very limited and expensive. Plus you are locked to one set of hardware. With a software solution you can switch hardware and not lose functionality. AMD is faster, go AMD, Intel is faster go Intel. Plus new features are way easier to add.
Cheers
Mike K