GTX Titan + GTX 780 GPU rendering?

[juliancrespi]

Hello!

 

I'm building a rig for 3d work mainly, planning to do a lot of gpu rendering with vray RT in 3dsmax.

 

I also been reading a lot this past weeks about different gpu options. After giving it a lot of thought I'm going for a GTX 780 for viewport, and a GTX Titan for Vray RT. Mainly due to the 6gb of the Titan, thinking next year I could buy another Titan and make them work in SLI, and with those 6gb it would work well for some years.

And for scenes with lower ram requierments I could use both the 780 and titan for rendering.

 

What would you think about these setup, any suggestions?

 

Thanks!

[Chris MacDonald]

I'm also looking at different options to, and I'm going to have 2x 4GB Nvidia 770's.

 

Nvidia 770 (~£400):

- cuda cores: 1536

- RAM: 4GB

 

Nvidia Titan (~£900):

- Cuda Cores: 2688

- RAM: 6GB

 

For less than the cost of a titan you can have two 770's, giving you more ram and more cuda cores.

[Dimitris Tolios]

I'm also looking at different options to, and I'm going to have 2x 4GB Nvidia 770's.

 

Nvidia 770 (~£400):

- cuda cores: 1536

- RAM: 4GB

 

Nvidia Titan (~£900):

- Cuda Cores: 2688

- RAM: 6GB

 

For less than the cost of a titan you can have two 770's, giving you more ram and more cuda cores.

 

Doesn't work like that.

 

Yes, 2x 770s have more cores than the Titan, and since we are talking "pretty much the same cores" (same SMX cluster / Kepler architecture) the performance increase is pretty linear and predictable. Cards "know" which is doing what, so things get evenly split between them.

 

BUT, each card is "on its own" as far as RAM goes, and needs to load 100% of the scene and assets independently.

 

　

 

Thus each 770 card will have up to 4GBs for exclusive use, the Titan will still have 6GB.

 

Same goes with SLI and gaming. Each card needs to do its own thing independently and then combine it, thus when companies advertise twin-gpu cards like the GTX 690 that is 2x GK104 with 2GB each as a "4GB card" or similarly the 7990 as a 2x3GB = 6GB card, they are simply misinforming their customers.

 

For GPU rendering, SLI is not required. Doesn't have any effect tbh, and as mentioned before, doesn't work with viewport acceleration within the 3D content creation apps we have today. You will be using only the primary card no-matter what.

 

Depending on your motherboad and ofc PSU, there are many combinations you could do if you can afford a 780+Titan right now, and get better overall results: if you have a 1000-1200W PSU for example and proper PCIe spacing/slots, you could go K2000 + 3x 770 4GB, or with an 850W K2000 + 2x 780s (depends on overclocking your cards and/or your CPU).

 

I don't know what you have in mind rendering, but 3GBs not dealing at all with accelerating viewport are pretty nice already: minimum 500MB saved there for just having a couple of monitors and 3DS loaded, without crazy stuff on. The most I have seen of allocated Vram when using 3DS/Maya, is 1.4~1.5GBs, so if a 4GB card is "ok" for doubling as a viewport and rendering accelerator, a 3GB one should suffice.

 

I have just gotten a used Titan myself, replacing my GTX 670 4GB. Wasn't about the VRam, just found it used for as much as I would pay for a 780 + tax. I have started testing, but did not work on it yet enough (will be going in a new built and I want to try 3DS Max 2014 at last). Quick results with OpenGL testing (Maya) saw the usual GeForce performance...it can be amazingly fast, but as soon as things get tesselated / complicated, it slows down a lot. In Maya it is still slower than Quadro 600 and much slower than the K2000 & Quadro 4000 using Specview perf 11. I will try rolling back drivers as the latest ones might have something weird, but unfortunately the Titan appears to be exactly where the 670 was at. No change, but with the drivers still I won't expect crazy changes (In Solidworks/ProE, the Titan is like 5 times slower than a $140 Quadro, not even 2 digit fps).

 

The 780 won't be faster than that, so as good as it is, with the current state of drivers/3D CAD API implementation, I would not recommend it as the "viewport accelerator" of choice should you have $650 (or more) to spent.

 

GPU accelerated renderings? Hell yeah! Innitial numbers with the Titan are nuts (Titan + 670 even better )

[Chris MacDonald]

Doesn't work like that.

 

Yes, 2x 770s have more cores than the Titan, and since we are talking "pretty much the same cores" (same SMX cluster / Kepler architecture) the performance increase is pretty linear and predictable. Cards "know" which is doing what, so things get evenly split between them.

 

BUT, each card is "on its own" as far as RAM goes, and needs to load 100% of the scene and assets independently.

 

That sucks so much. Maybe I'll re-think my investment and get a k2000 or k4000.

[Dimitris Tolios]

That sucks so much. Maybe I'll re-think my investment and get a k2000 or k4000.

 

Yeah, well...it is what it is (at the moment). HSA might change that, promises common pool for CPU + GPU for RAM and resources in general. But remember, its AMD that announced its implementation, so if things keep being biased towards nVidia (and AMD CPU side doesn't do more vs. Intel in the high end segment), it will take quite some time for things to be really different. Definitely not before you complete this built, probably not even before it is time for it to be upgraded for other reasons

 

It is sad to know that all this hype for GK110 (which I own, even sadder) and its uber performance, is actually applied in games only (and folding, where the Titan does 3-4x times what my overclocked 3930K does, using 2/3s the power).

 

In the CG world, it is so simply because nVidia plays in its own game.

OpenCL compute, although amazingly faster than the 670 (and 680/770), it is sad to know a 7950 would be even faster for 1/4 the cost (actually 1/5th now that retailers are selling out for the next generation of Radeons, and 7950s in the US are dropping below $200).

 

Anyways, what I've said above still stands (for multi-GPU configs), but I would suggest you pushing forward after testing the waters 1st.

Make sure you know what VRay RT GPU can and cannot do for you before you go "all-out"...

i.e. a single 770/780 or w/e will work for you to see what it does and what it doesn't, what you can adapt to and what might be a deal breaker for your workflow. Getting a 1000-1200W PSU that you might end up not using, is a big expence, but getting $1500 or more worth of cards that you will end up not using is much worse.

[Chris MacDonald]

Might just go with a K2000 and a decent PSU, then in a few months see what is worth getting.