GPU rendering is not the magic bullet it was promised to be

Thu Aug 12, 2010 11:16 am

radiance wrote:If i compare a quad core CPU doing 0.3 megasamples/sec against a GTX480 doing 4 megasamples/sec

Where are these numbers from? If this is true then Octane on a 480 would have to be 13 times as fast as Maxwell. Nothing at all points towards this being even remotely the case.

Thu Aug 12, 2010 11:52 am

Are you disappointed with Octane Chris_TC?
Would you like your money back?
Everyone is happy here except you for some reason.

Thu Aug 12, 2010 12:03 pm

pixelrush wrote:Are you disappointed with Octane Chris_TC?
Would you like your money back?
Everyone is happy here except you for some reason.

I think Chris is just being reasonable actually. In my test that I linked to earlier Octane on a GTX470 is about 25% faster than Maxwell on an i7 980X - so extrapolate that to a quad-core CPU vs a GTX480 and you might get to 50% faster. This is on a very simple scene, once the scenes get more complex it's harder to judge because of the fireflies, but once Octane has MLT I will test again. I don't see a 10x increase in speed, and I don't think there's anything wrong with pointing that out.

Maybe in the future there will be optimisations that will give a speed boost, and of course Octane is still in Beta so it deserves some slack, but nonetheless I don't see a problem with just telling the truth.

radiance · Thu Aug 12, 2010 12:13 pm

Chris_TC wrote:
radiance wrote:If i compare a quad core CPU doing 0.3 megasamples/sec against a GTX480 doing 4 megasamples/sec
Where are these numbers from? If this is true then Octane on a 480 would have to be 13 times as fast as Maxwell. Nothing at all points towards this being even remotely the case.

Hi,

On a sample by sample basis it is, but maxwell uses bidirectional path tracing and MLT, which makes it converge faster.
Maxwell is now what 5-6 years old ? You can't compare an advanced, efficient bidir/MLT renderer with a bruteforce path tracer.
maxwell can get away with 2000 pixel samples for a clean image with lots of glass and caustics.
octane needs maybe 64000 due to the missing bidir/MLT.
however, octane can do ~30000 samples in roughly the same time maxwell does 2000.
So once octane needs 2000 samples, when MLT and ultimately bidir are added, you will have a 10-15x speed-up

Radiance

Thu Aug 12, 2010 12:25 pm

radiance wrote: So once octane needs 2000 samples, when MLT and ultimately bidir are added, you will have a 10-15x speed-up

Radiance

Now that is something to look forward to

I really hope that Octane can do this at some point in the near(ish) future!

Thu Aug 12, 2010 12:36 pm

I think Chris_TC is being something of a worry wart.

I note that there is a lot of technical nitpicking going on in that thread as if Octane is on a development par with Maxwell or Modo which it isnt.
I resent the implication that Refractive are apparently not telling the truth although it is not my business to defend it.
As time goes by this seems more and more like a smear campaign come to spoil our nice 2.3 party.

Thu Aug 12, 2010 12:40 pm

Pixelrush: don't be so paranoid

No one is trying to "spoil" Octane. My motivation for doing my test was simply to compare the capabilities of three different rendering solutions, with no preconceptions as to which would be "better". If Octane lives up to its promise I will be delighted.

Thu Aug 12, 2010 12:49 pm

pixelrush wrote:Are you disappointed with Octane Chris_TC?
Would you like your money back?
Everyone is happy here except you for some reason.

Would you stop being so defensive and taking the blinders off? I obviously don't want my money back, neither am I convinced that CPU renderers can keep up with Octane speed-wise. But I at least consider the possibility of this being the case when several independent tests indicate it.

radiance wrote:So once octane needs 2000 samples, when MLT and ultimately bidir are added, you will have a 10-15x speed-up

Okay, thank you for clearing it up.

Thu Aug 12, 2010 12:56 pm

pixelrush wrote:Are you disappointed with Octane Chris_TC?
Would you like your money back?
Everyone is happy here except you for some reason.

Hi Pixelrush,

I did not know that you work for refractive. Why on this earth would you else be so quick in giving people their money back and drive them away from Octane. Do you think you do Refractive a good service when your last argument (now in several discussions) is that critical people should quit using Octane. It is not your enterprise so please do Radiance a favour (at least in my personal opinion) and stop telling people to get their money back and leave because they might do it and thereby harm octane (and it is much to good a piece of software that this should happen).

Lutze

Thu Aug 12, 2010 1:08 pm

People wouldnt be pursuing gpu computing if there werent decent gains to be had.
Its not imaginary - go take a look at the nvidia cuda site. http://www.nvidia.com/object/cuda_home_new.html
There are many different applications there with a large speed up for calcs.
The dude in the Lux video muttered something about 10-100x faster and I have never heard anyone claim Octane gave that.
I think you guys should cut Octane beta some slack and stop fretting and obsessing about early performance.
Here is Furryball http://furryball.aaa-studio.eu/ making claims of 30-300x faster on gpu - seems a little extravagant - perhaps you could benchmark and compare that one instead if you want to worry about something.

Lutze <groan>