Entertainment, Graphics, Hardware, Software Programs

The Nvidia Kepler Render Review ? Is Titan King?

The GeForce GTX 780 and Titan cards from Nvidia are the cards that left me with a big surprise from the moment I opened the box. How is this possible?

If you compared these cards to the car industry and the most wanted automobiles, these cards would be a Lamborghini Gallardo and all other Nvidia cards would go into the class of the Volkswagen Golf, that is, design wise. This is literally one of the best looking designs of graphic cards I have seen so far. It’s mostly made of metal parts, and is so nicely built that I wish we had more of these units we could put on the shelves in our offices. The lines are nice and smooth. And the grill, covered with the transparent plastic, is just like the glass on the back of the Lambo covering the engine, suggesting there is more than meets the eye. 

While looks are not crucial for decision making, when buying graphic cards, I just needed to take that thought off my mind. The cards we had great pleasure of reviewing had a great solution for cooling the hardware. They take the cold air from within the chassis and blow it the hot air out of it. This solution is especially great for the vertical cases (Like Maingear?s SHIFT) where all of the Quad-SLI cards would exhaust the warm air out the top and leave the inside air, cold and optimal in temperature.

Keep in mind, we?ve already reviewed the GeForce GTX 780 and The GeForce GTX Titan, but only from a gaming and architectural standpoint. As you could tell from our reviews, they did pretty well in gaming, however, we were left wondering how they would fare in professional applications.

With all of the standard connections available this cards don?t fall short from being just great. Or do they?
Doing tests in a real life production environment, we stumbled across some interesting numbers that don?t make much sense, which can only be explained by the fact that these cards are so new that they need to give developers more time to optimize their software for these ?new players?.

As we did in our GeForce GTX 680 review, we wanted to see how well these new cards behaved where we needed them the most; and that is in the GPU rendering of a 3d scene. (Mostly because we found our best bet in a CAD/CAM environment with the Nvidia Maximus 2 configuration with Quadro K5000 and Tesla K20c, and there is not much to add at this moment).

As you might remember from our last No BS article (the one with the Maximus 2 configuration) we realized that the best way to go is to setup our production units as two individual machines. One for content creation and the other one for GPU rendering. This article is trying to get some answers on the part of a GPU rendering machine configuration.

The GTX 680 machine gave us some nice numbers and we did ?beat? the ?Maximus Rig? in GPU rendering but we also noticed there was some room for further development of things. Renderings done on Open CL engine gave us results we couldn?t quite figure out and we wanted to what the truth is, utilizing new generation of GTXs and the problem continues. Open CL is just not as ?fast? as CUDA, at least over Nvidia based cards, how it behaves on the last generation of AMD cards, remains to be seen, but at this moment the numbers are as follows:

Our conclusions can very easily be drawn from this table. CUDA performance suggests that the new cards gave us better results compared to those of the GTX 680s, a year or so old cards. This is not being concluded from some synthetic tests made to ?make-or-break? the card/manufacturer, but rather real life tests made in Vray RT ver. 2.4. The performance scales almost linearly through the generations and the models (GTX 680 to GTX780 and then to Titan, leaving room for GTX 690 to fit somewhere on the ?straight? line of performance chart).

Since we already had the scene made from the last tests we made with Maximus, we decided to spice up things a bit. We added some samples to be computed, bigger textures, and some polys just to make scene more ?compute heavy? so we can investigate the stability and seriousness of the rig under greater pressure, as we did last time with the computation of a animation. No lags, bugs nor instabilities were found.

There are several things you can read out of this table:

Titan ? The Second Generation Kepler
As you can see, Titan card compared with the GTX 680, will give you 40% better results in GPU rendering than GTX 680 and having 4 of those in the Quad-SLI configuration will make your renders explode (in a good way).

Titan vs. GTX 780
This was the little disappointment we had stumbled upon. When we were looking at the specs of both of these cards, we were hoping that the higher clock speed on the GTX 780 will give us slightly different results than Titan, but the fact that is 35% slower than Titan made us think about recommending to spend the extra $350 on a Titan (Notice the correlation between 35% of better performance and the difference in pricing, $350 – Prices compared using Newegg.com?s price list). But then again, there is no free lunch these days, except if you are over at you mother?s place. Unfortunately we haven?t had a chance to test 4 Titans against 4 GTX 780s to make more sense in overall gain of performance, but you can do your own math until BSN brings you this numbers first hand. 

OPEN CL ? Enigma or strategic move on Nvidia?s part?
Looking at these numbers for Open CL rendering is just amazing. Either the Chaos Group didn?t have the time to optimize for Open CL on the new generation of Nvidia cards, or there is a long term plan on Nvidia?s part to make CUDA the ?way to go?, attempting to render AMD obsolete. Making hardware accelerated software that is used by developers in their software (ex. Rayfire, demolition plug-in for 3ds max) that is ?optimized? for Nvidia hardware. Not to mention that they have pulled companies like OTOY completely in, and helped them develop one of the best renderers on the market only CUDA based. This is in addition to the big production companies that have switched to Nvidia?s high end solutions, but then again, we are maybe jumping ahead of ourselves.

Even if Chaos Group, the maker of Vray rendering engine, optimized for new cards, it is highly unlikely they will be able to bring down the render times so much that it would actually be useful compared to CUDA based cards. Is AMD losing the API battle in GPU rendering? Well, we will see, but something funky is going on here.

The fact is that Titan cards are the predecessors to the upcoming, new range of profession
al Quadro K6000 cards, tells you the most about these cards. They seem to be just like the scouts spawned on to the enemy?s territory to see, from the distance, what would be the best impact of an upcoming storm. The GTX 780 is a little brother to a big Titan. Not to be neglected and forgotten but its full potential is clouded by an overprotective powerful brother.

We’re hoping to be able to test AMD’s new R9 290X against Nvidia’s Titan in OpenCL to get an accurate and up to date measurement of who’s faster in rendering. Hopefully that happens sooner rather than later. And hopefully clear up the whole Vray Open CL question.