Editorial NVIDIA GeForce GTX TITAN Finally Benchmarked and Rated!

By now, many or most PC enthusiasts will have seen the reviews of the NVIDIA GeForce GTX TITAN graphics card. I'm not going to rehash everything that those reviews have revealed about this card, but I will talk about it and link to the reviews that are in my opinion, the best ones. If you would like to read about the various features and technical details of the card, have a look at the reviews below, or check out my overview, here.

First off, the price: at launch, it costs £830 / $1000 which is just about the same or a little over the current price of a GTX 690 and is of course seriously expensive. This is interesting in a way that I'll come to soon. Of course, the price is so high, because AMD doesn't have a product in this performance class and hence no competition, making for a similar situation to its CPUs lacking a top end part to compete with Intel's top end.

Then we have the physical construction of the card. It's clear that it's been given the same treatment as the GTX 690 which precedes it. This means it has a very similar look and is built to a very high standard with the overall quality of its engineering being top notch. It also features that smart, distinctive magnesium unibody construction.


GK110 GPU die shot

Then we have the whole point of its existence: performance and lots of it. With its 7.1 billion transistor Kepler GK110 GPU and massive 6GB GDDR5 memory, it's very fast indeed and can play any game you throw at it maxed out and at high resolutions as the benchmarks have revealed. It also has two interesting new features which help to wring the maximum performance and gameplaying experience out of it: GPU Boost 2.0 and display overclocking.

So, why is the price interesting? Because its a bit of a swings and roundabouts situation when compared to its direct competitor, the GTX 690, leaving no clear winner when everything is considered. In fact, I'd go as far as to say that only one of these cards is going to survive in the market, but I don't think we have a clear winner.

First off, the performance is disappointingly actually somewhat lower than a GTX 690, even when overclocked to a higher GPU and memory clock speed, as the results below from TechPowerUp's review show:



In the above test, the TITAN's GPU was overclocked from its stock speeds of 837 / 6008MHz to 1005MHz / 7020MHz which compares to GTX 690 stock clocks of 915 / 6008MHz. Note that the memory speeds here are effective, since GDDR5 is quad pumped. This means that a 1502MHz real clock gives an effective clock speed of 6008MHz, quadrupling the data throughput of the memory.

As you can see, the GTX 690 at its stock speed is still around 8% faster, which to me, is disappointing. I guess the reason for this result is that the dual GK104 on the GTX 690 effectively makes for a "bigger" GPU since it has an effective bus width of 512-bit and 64 ROPs when one considers the two GPUs as one chip. This compares to a 384-bit bus width and ROP count of 48 for the TITAN.


Block diagram showing one SMX unit disabled
in the GK110 GPU fitted to the TITAN..
Click to enlarge.

Remember that this consumer version of the GK110 GPU has one SMX unit disabled making it a crippled chip, however slight, as the block diagram above shows. Hopefully, in time NVIDIA will release an upgraded version of the card with the full uncut GK110 GPU in a similar way that the GTX 480 with its one disabled cluster turned into the GTX 580 which enabled it and improved various aspects of the GPU, producing a significantly improved graphics card once manufacturing yields improve. On the up side, it looks like the TITAN will be a good overclocker, however.

Then we have other factors to consider.


The Memory

The TITAN has 6GB while the GTX 690 has 2 x 2GB. The 3x more available framebuffer memory is likely to make a significant difference in multi-monitor scenarios or 4K resolution (4096 x 2160) where the GTX 690 is likely to run out of memory, tanking performance and I really mean tanking. The minute data needs to be pulled over the PCI-E bus rather than residing in the frame buffer, performance drops off a cliff. This won't happen with the TITAN. Then again, how much performance is left in the GPU at these monster resolutions? Not that much judging by TPU's benchmarks:



For me, it's 60Hz minimum with vsync on or bust. In fact, with my Asus VG278HE monitor running at 120Hz refresh and LightBoost eliminating horrible LCD motion blur, I'm shooting for 120Hz minimum.

After all, what's the point of looking at a beautiful picture in your first person shooter when it's juddering and hitching all over the place? None. It totally ruins the gameplay experience. Note that these results were achieved by maxing out the quality settings on demanding games, too. Therefore, it's quite possible that running older games or more modern ones at more conservative settings will significantly alleviate low framerates. Unfortunately, I don't have such a rig to try this out for myself and report on here.

Finally, 6GB of GDDR5 is still expensive, so a cheaper 3GB version of the TITAN would surely be just as good for the majority of gamers out there and save perhaps £100-£150 on the price? Somehow I doubt that NVIDIA will do this, since this card is more about having the crowning glory in single GPU performance rather than making any serious cash directly out of it. Or in other words, marketing. This is important for boosting sales of a whole lot more lower end cards which is where the real cash is to be made.


SLI Setups

Two GTX 690 cards can be run in SLI, making for a 4 GPU setup delivering awesome performance. However, the TITAN can also be run in a four card configuration (time to for a bank loan...) also making for a 4 GPU setup. Now, for very well-heeled enthusiasts, the performance tables between the two cards are now turned. In this configuration, you have one GK110 going up against one GK104 GPU and the result of course is that it obliterates the GK104. Therefore, if money is no object, then 4 TITANs are the way to go! A quick calculation works out to around £3,360 for all four TITAN cards...

In fact, just three TITAN cards will significantly outperform two GTX 690s as the benchmarks below from TPU demonstrate:



Of course there's a couple of well known caveats with multi-GPU setups, regardless of whether it's two GPUs on one card or on different cards: scaling and memory availability. Unfortunately, doubling the number of GPUs on the card does not double the performance, due to overheads such as the necessary communication and data that needs to flow between them to make them work in tandem to render the one scene. The latency this introduces is also a significant factor, which leads to greater microstutter where a moving scene can appear juddery. Note that this happens to single GPU cards too, but to a smaller degree, as there is no latency issue between GPUs.

Scaling can be an issue too, with optimized SLI profiles for each game being required to the most from the card. Hence, a poorly supported game may run as badly as only harnessing the power of one GPU rather than both. That's a bit of an extreme situation, though.

Finally, there's the memory available as frame buffer for the game. Having 6GB available on the TITAN means that it can use all of it for storing scene data and general processing. However, with multiple GPUs each one must have its own framebuffer, which effectively reduces its size. Therefore the 4GB on the GTX 690 is really only 2GB available for the game, since each GPU must have its own dedicated memory buffer, duplicating the scene. This leads to that memory runout problem at high resolutions that I explained above - the very time you'd want to have as much memory as possible to work with.

Also, it's now easy to see how memory use becomes increasingly inefficient as the number of GPUs goes up. So, for a two card GTX 690 setup, we have 8GB RAM of which only 2GB can effectively be used, "losing" 6GB. For a four card TITAN setup we have 4 x 6GB which equals a gargantuan 24GB of RAM. Yet, the game sees only 6GB, "losing" us a whopping 18GB! Of course, 6GB is still massive by current standards making the "loss" of all this RAM a relatively moot point for now. Still, the point is made on the inherent limitation of using multiple GPUs together. I will write another editorial soon on a theoretical way to run multiple GPUs in tandem which eliminates all these problems!



Conclusion

So, NVIDIA's single GPU flagship graphics card, the GeForce GTX TITAN has now been fully unveiled and benchmarked and unsurprisingly, it's a really fabulous card, easily taking the single GPU performance crown. And yes, I want it, very much. I won't spend that kind of money though despite having a good enough rig to put it in as there are much more important things in life to spend that kind of money on. Oh, what the heck, I'll preorder one now. No, kidding. The card has been paper launched for now, with general availability slated for February 25, which most likely translates to something like 10 cards for the whole country...

One great thing about this card are its quiet noise levels, better than the GTX 690 even, since impressive graphical horsepower while driving you mad with the noise is nothing. This is an area where AMD has to seriously improve.

So, which card do you think is better? Which one would you buy? Which one do you think will live on or do you think both will? Do GPU Boost 2.0 and display overclocking really do it for you? Let us know in the comments.

I'd like to thank TechPowerUp for providing all the pictures and graphics for this article.

And now for those links:

TechPowerUp (single card)

TechPowerUp (three card SLI)

HEXUS

Bit-Tech

PC Per

HardOCP

ExtremeTech (This is actually an interesting article on what the TITAN graphics card means for the future of PC gaming and AMD)