OTOY Forums

Hi all,

I saw some benchmarks results here in the forum and I noticed that my HW solution is very slow when compare to a single GTX 480.

When I use both cards (Quadro 5800 + Tesla C2050) and I load the famous benchmark scene, in pathtracing mode I get 3.9 Megasamples! I noticed that a single 480 produces 4 to almost 4.5 Megasamples!!!

What happens here???

I should produce almost the double performance of a single 480 card...

Do you have any idea? I use the 2.2 beta version and I also get the same results with the pre 2.3.

Thanks in advance,

Dionisio -

Hi,
Your Tesla 2050 have 448 CUDA processors and this puts it around GTX 470 performance.. and with combination of a slower card, in this case Quadro FX 5800 with its 240 CUDA processors, witch correspond to GTX 275 performance, it's pretty normal to experience that computing slow down...

Thanks for replying.

The quadro 5000 is based on Fermi technology and it has 352 Cuda cores.
So in theory a Tesla C2050 + Quadro 5000 ( both Fermi Gpu's) can't be lower than a single 480!!!
It sounds very strange to me.

We're planning to build a 16 GPU renderfarm and we're testing some cards. The problem with the 480 is the heating and memory issue as 1.5 GB is not too much for professional productions...

Dionisio -

Hi, here is the link to specifications of the Nvidia FX 5800 card : http://www.nvidia.com/object/product_qu ... 00_us.html .
It's clear that this model is NOT fermi based and have only 240 CUDA processors.
However, this model here : http://www.nvidia.com/object/product-qu ... 00-us.html IS fermi based, but the name is QUADRO 5000 without FX..
So the model in the thread name (Quadro 5800) is NOT femi based and have 240 CUDAS and it will slow down your TESLA card calculations..

The new QUADRO FERMI based card are : Quadro 4000, 5000, 6000 and the FX in the name is removed..
Here is the link to the quadro cards available: http://www.nvidia.com/page/quadrofx_family.html
Notice: newer models are marked with NEW! after their names, meaning FERMI cards..

So, Octane performance linearly scales with CUDA cores, only within the same family of processors?

That was my error! It's a Quadro 5000 (Fermi series).

I changed the title of my post.

So this is what I am saying, having 2 Fermi boards, even if individually are slower than a 480, together should be faster and not slower than a single GTX 480 board...

Do we have any possibility to check this "bad" performance?

It's an Octane's problem that it doesn't fully support Fermi cards yet?

Thanks,

Dionisio -

Hi,

You should do tests on a heavy scene.
If you run the standard benchmark scene, the cuda kernels do the passes so quickly they are slowed down by the tonemapper limit.
therefore, you will notice the speedup once you turn on pathtracing, and use a lager render resoliution, like 3000x2000 or equiv.

Radiance