Multiple GPU vs dual Xeon system?

[reyner]

How many titans cards will be needed to match the rendering times of a dual Xeon system (E5 2640-2.5 GHz) with a clean result?

[tricky]

Isn't GPU rendering like a zillion times faster than a CPU? (its just that certain features aren't supported yet)

[Macker]

It's a valid question. The only thing currently stopping me from recommending the practice I work for to switch to GPU rendering is the amount of RAM onboard the gpu. We'd need 12gb minimum really. Everything else that we use in every day practice is supported by RT now, even forest pro.

[vlado]

We'd need 12gb minimum really.

You might actually get that sooner than you think

Everything else that we use in every day practice is supported by RT now, even forest pro.

Very good to hear it

To go back to the original question, it very much depends on the particular scene - i.e. interior/exterior; image resolution; material complexity etc. For many things, especially exteriors, the Titan is insanely fast (compared to a brute force approach on the CPU of course).

Best regards,
Vlado

[tricky]

Everything else that we use in every day practice is supported by RT now, even forest pro.

But not MultiTexture or BerconTile...c'mon guys. Who do I need to pester? Mr Bercon? Mr MultiTexture? Are there 'similar' tools available that are supported by RT?

[reyner]

Hi, Vlado, thanks for your input . What I´d like to know is this: How many Titans cards will be needed to match the rendering times obtained with the biased combination of irr.map/light cache on dual Xeon E2640. I know it may be difficult to estimate because of different approaches to solve the GI but even a rough idea will be great. We do mostly office interior renderings @ 3500px by 2000px.

[vlado]

Well, get me a scene set up with the production V-Ray to vlado@chaosgroup.com and I will try to run some tests and see how it goes. We can put a number of cards and CPUs to it and see how it goes.

Best regards,
Vlado

[Macker]

I shall be watching this thread with bated breath, eager to see some results.

[Macker]

But not MultiTexture or BerconTile...c'mon guys. Who do I need to pester? Mr Bercon? Mr MultiTexture? Are there 'similar' tools available that are supported by RT?

Ahh, I forgot about multitexture. That is pretty handy =/

[Zach Gray]

You might actually get that sooner than you think
...

Best regards,
Vlado

Hmmmm.... a hint at 'out of core architecture' like redshift?

That would be cool.

[eyepiz]

That would be cool...Redshift looks very interesting.

Hmmmm.... a hint at 'out of core architecture' like redshift?

That would be cool.

[Alan Iglesias]

Isn't GPU rendering like a zillion times faster than a CPU?

At basic path tracing, yes. But often, especially in the case of interiors with plenty of bouncing light, getting an acceptable image in terms of noise can require enough path tracing to send the number of paths traced per pixel through the roof - so many that even a multi-core GPU setup can take some time to get there.

Note that in the kitchen curtain scene I posted, the number of paths traced per pixel was a whopping 12,040 - many, many more than most images I render with RT/GPU. In these cases, greatly reducing the amount of paths to trace using biasing techniques can reduce noise in a very fast, efficient way whereas our non-biased GPU renderer still renders every single ray in the scene to get a solution.

At the current time, GPUs are just getting strong enough to begin to challenge the CPU/biased renderer for interior shots with lots of bouncing GI to compute. In this thread's case, it would be good to get a typical shot rendered by the OP and test it in GPU and the CPU/biased render with a specifc CPU(s) to truly see where he is at with the technology, based on his personal needs. He is in the same basic place as you, Tricky. It would be interesting to see the results of such a test...

-Alan