2.4 pre-beta speed improvement on 460 gtx

Mon Jan 17, 2011 3:11 pm

In our test speed improvment with cuda 3.2 get 460 gtx 15-20 %.
I think, it is not enough.
Gtx 460 has 1/3 more cores and also radiance was talking about 15-20 % improvement with cuda 3.2 against to 3.0.

So I was expecting 70 % (50 + 20) better performance in this version.

What is wrong ?

Mon Jan 17, 2011 3:44 pm

My improvment in my tests is about 20% too.

Mon Jan 17, 2011 4:35 pm

Yep i have the same on my 460 as well about 10% speed improvement which i suspect is down to the optimising of the code on this release rather than actually using the other 112 cores.

One thing has occured to me.

What if Cuda was using all the cores all along but was just reporting that it saw only 2/3rds of the cores?

i wonder if there is any way of checking?

Just a thought....I might be wrong.

Mon Jan 17, 2011 6:07 pm

Jaberwocky wrote:Yep i have the same on my 460 as well about 10% speed improvement which i suspect is down to the optimising of the code on this release rather than actually using the other 112 cores.

One thing has occured to me.

What if Cuda was using all the cores all along but was just reporting that it saw only 2/3rds of the cores?

i wonder if there is any way of checking?

Just a thought....I might be wrong.

disabling alphashadows increase speed of another 10 %

Mon Jan 17, 2011 6:42 pm

Seems to be like only Cuda 3.2 speed increase,
maybe 336 cores is only showing but doesnt working

Mon Jan 17, 2011 6:59 pm

You have the tool that's shipped with Cuda toolkit and it's called "Visual Profiler".... Or the simplest way to monitor GPU cores usage is GPU-Z ver 0.5, hope this helps.

If you check the graph in GPU Z you'll see that with Cuda 3.2 (Octane 2.4 ver) cores are constantly on 95%, wheres on 2.3v5 there's oscillation in the cores usage. This is (i believe:) where the difference in speed is coming from. Sorry for my English.

All the best,
Voja

Mon Jan 17, 2011 7:07 pm

Unfortunately, there is more to speed than only the number of cores. All these additional cores are only of any help, if you can utilitze them, i.e. the code runs fully parallel. How parallel things can be executed, depends on the geometry, materials and kernel settings and other stuff. -> An increase of 50% render speed is not possible and won't happen.

Also make sure that you disable alpha shadows if you want to compare render speed.

Cheers,
Marcus

Mon Jan 17, 2011 7:16 pm

Hi Marcus,

On my system the speed is the same with or without alpha shadows - literally identical. Again, that's on gtx 580. I just wanted to point out the tools for monitoring GPU usage. Thanks for your explanation and input on the subject.

Regards,
Voja

n1k · Mon Jan 17, 2011 7:33 pm

Well, U should have heavy use of alpha mapped geometry to see difference between alpha shadows on/off. Try rendering tree, for example, and speed impact should be obvious.

Cheers,
n1k

Mon Jan 17, 2011 7:36 pm

infernoVFX wrote:Hi Marcus,

On my system the speed is the same with or without alpha shadows - literally identical. Again, that's on gtx 580. I just wanted to point out the tools for monitoring GPU usage. Thanks for your explanation and input on the subject.

Regards,
Voja

Yup, thanks for that

Regarding the speeds with and without alpha-shadows: If you use only HDRI lighting and no emitters and no sun, then the alpha-shadow option becomes indeed completely unimportant.

Cheers,
Marcus