GPU test

[benfattino]

I try to do some test with different configuration.
The scene is the same for every machine.

GTX570 2,5 GB - 550 sample/s - pathtracing
GTX660 2 GB - 700 sample/s - pathtracing
GTX680 2GB - 910 sample/s - pathtracing

MacPro Dual XEON 5520 - 900 sample/s - bidirectional pathtracing

I tried to create a scene with elements of all kinds. (emitter, displacement, metal, glass)
This is just a test to understand the differences between the cards nvidia Series 5 and Series 6.
It would be interesting to do a test with an ATI card (possibly 7970 6 gb) or Tesla (maybe k20G).
There remains the question: should the gtx680 than the GTX580 (among other things now lost)?
Small note: a MacPro with Dual Xeon still behaves very well (although of course the cost is much higher than a single card) ...

I hope I have implemented a good test. I accept tips ...
S.

[Zom-B]

Ok here some advice to get some more performance out of GPU rendering:

- Set the splat filter to fastbox
- manually set the threads for Indigo to CPU Core count with HT +1

Also GPUs starts flexing muscles in exterior scenes or Studio setups where rays can be killed quickly by bouncing off into space instead of going through a high ray depth in a closed room.

In the end the problem is always to get good balance between CPU and GPU power so noone of both starts to idle...

[benfattino]

Thanks for the tips.
I only understand if 680 is a good choice... And I think it is...
If i use 4 MacPro on network render (bidirectional pathtracing) the result are more higher...
Hope someone test ATI or Tesla...

S.

[Zom-B]

The thing is that Indigo uses hybrid rendering, CPU + GPU together, so the harmony between both components is important, otherwise one of both is overpowered and idles! The overall speed gain from rendering together with a GPU is not super-ultra-mega-high, but +100% and more can be achieved in certain situations.
Test your PC without GPU but still with pure pathtracing, so you'll see how much benefit you get...

BiDir PT in general generates less samples per second, but converts complicated situations faster to a clean image the pure PT (caustics for example).

[Mor4us]

so why is there no gpu only rendering if in most scenes gpu is limited by cpu speed/ you're always having a bottleneck between cpu and gpu conversion.
fe for people who can't afford a dual xeon monster workstation there is no way to gain full advantage of buying a new gpu as long as your cpu is too slow to feed it... :-/

[Zom-B]

so why is there no gpu only rendering if in most scenes gpu is limited by cpu speed

Not most scenes, but Interiors mostly

in the hybrid approach the GPU is only doing a part of the work, the part that suits the GPU best.
This approach saves a lot of RAM on the GPU and was great in times it hit the market, when the top RAM amount a Graphic card had was 1GB.
Also in the time Indigos GPU rendering was developed the whole situation was kind of new and undiscovered,
porting a whole engine to GPU was kind of a big deal these days.

I have no idea what Indigo future plans here are and where that journey is heading too, but I can tell that in the Studio setup animation of a couch I'm doing here atm for a customer my shity Nvidia GTS 450 bumps the samples per sec by +80% on my 6 core AMD 1080T @ 3.5Ghz and rendering in pure PT is due to the "simple" materials of a couch a winner regarding quality over time compared to BiDir and MLT.

[zeitmeister]

- manually set the threads for Indigo to CPU Core count with HT +1

Erm, can you please explain that again? Or in german?

[Zom-B]

- manually set the threads for Indigo to CPU Core count with HT +1

Erm, can you please explain that again? Or in german?

I discovered that Indigo renders faster if you have manually set it to render with one extra thread compared to the auto Threads Indigo uses.