Geforce Titan / 780GTX

[itou31]

Nice !
It is done with all CUDA parameters active ? (GPU Boost OFF).
Is octane based on FP32 or FP64 ?

[face]

Nice !
It is done with all CUDA parameters active ? (GPU Boost OFF).
Is octane based on FP32 or FP64 ?

FP32 so much i know...

face

[JuM]

My priority is not the speed but the memory. I hope they make 8 or 12 gb versions of it

[p3taoctane]

I see somewhere in this thread someone saying you can't run more than 8 GPU's in one system.

My experience backs that up as I tried to get 9 going once with 8 rendering and one display. (Using an extension box) and it only saw 8 cards.

It is a little off topic... but WHY IS THIS. Is it a windows limitation, a mother board limitation?

There are Trenton Expansion chassis with 17 PCIe x 16 Gen 2.0 slots so you could fit 8 doubles at least with a 2430 watt PSU incorporated in it. Then as many as 7 more in your connect CPU

I'm guessing if this could have been done... someone would have done it. I tried working with a company to do some tests on cloud rendering... but they were only letting me log into 4 card systems. If OTOy wants to do cloud based rendering I am assuming they will be doing more that 8 cards.

I do not plan on going out and purchasing 16 Titan cards anytime soon... but I bet 580s aree going to come down eventually and a 16 card system would not be out of the question.

Kind of a dreamer topic... but has anyone here looked into why it is only 8?

Peter

[darkline]

Now that the performance review ban has been lifted: We don't have any cards here yet, but Chris gave it a spin with version 1.10 on one of the review site systems via remote access and the results were, that the Titan is on the benchmark scene (maxdepth 16, no alpha shadows) 60% faster than a GTX 580 and about 10-12% slower than a GTX 690 (using both GPUs).

In my opinion that's fairly impressive and better than I expected and although expensive it may be a good GPU for smaller studios that actually earn money with Octane. You basically get almost the speed of a GTX 690 with 3x as much memory. We have to wait now until those graphics cards actually get sold and arrive here in NZ...

Regarding "optmization" for compute model 3.5 GPUs: A while ago I tried to coax more speed out of a K20c using different build options, but it turned out that the way how we currently build the CUDA stuff for Octane, was already the fastest. I will have a look at it again, when we hold a Titan in our hands, but don't hold your breath. There may no speed gains here.

Cheers,
Marcus

Marcus,

This is exactly the feedback I was looking for - thank you. 60% faster than a 580 is a good starting point even if you cant squeeze the GK110 for any more. Keep us updated with any optimisations as this would really tempt me to buy one over a gtx690

thanks again

[MOSFET]

Now that the performance review ban has been lifted: We don't have any cards here yet, but Chris gave it a spin with version 1.10 on one of the review site systems via remote access and the results were, that the Titan is on the benchmark scene (maxdepth 16, no alpha shadows) 60% faster than a GTX 580 and about 10-12% slower than a GTX 690 (using both GPUs).

Hey Marcus,

I asked a review site earlier with 3 x Titans if they could run the benchmark scene using the demo version (maxdepth 8, alpha shadows enabled) and they were kind enough to give it a go:

1 x Titan = 6.25 Ms/sec
2 x Titan = 12.50 Ms/sec
3 x Titan = 18.86 Ms/sec

While Fooze's 4 x 690 setup at 27.77 Ms/sec should still beat a quad Titan configuration (~25 Ms/sec), having 6GB vs 2GB of VRAM is very tempting.

-M

[gabrielefx]

4 x Titan = 25M

less noise
less power
more ram
more textures

[itou31]

Thanks for theses bench.
my 2x580 1.5GB do 8.80 Ms/sec (maxdep 8, alphashadow), but often lack of memory.
I'm searching for 580GTX 3G, if cannot, perhaps a Titan soon (for its 6Go and possibility to put it in my little tower)

[mbetke]

itou31, I will have two 580 3GB for sale soon. ,)

[mainframefx]

My priority is not the speed but the memory. I hope they make 8 or 12 gb versions of it

Wont happen, manufactures are bound to nvidias standard design.