NVIDIA GeForce GTX Titan - The King of Kepler Tested
Sharing similar specs to the Tesla K20 enterprise part, the GTX Titan boasts 2688 CUDA Cores and 7.1 billion transistors, making it by far the most powerful single GPU ever created for consumers. We pit this FPS crunching graphics card against the best setups possible to find out if it deserves to demand US$999.
By HardwareZone Team -
Note: If you've read our introductory article on the GeForce GTX Titan a few days go, you can skip this page and jump straight into our Test Setup and Performance pages.
A Brief History of the GK110 GPU
NVIDIA's GK110 die
Last year, in March 2012, NVIDIA unleashed its 28nm Kepler-based architecture on the world with the release of its flagship Kepler GK104 GPU-powered NVIDIA GeForce GTX 680. And while it was, without a doubt, one of the most powerful GPUs we had seen to date, not everyone was satisfied. You see, while GK104 was good, it wasn't GK110.
Enthusiasts who had been following the progress of Kepler pointed out that GK104's numeral code was an indication of a GPU in the middle of the range, not a flagship product. Speculation was rife that an even more powerful GPU existed, a yet to be revealed GK100/GK110 GPU that would hold the true power of Kepler. Rumors of its specifications and capabilities were whispered, and a synopsis for NVIDIA's (then) upcoming Graphics Technology Conference 2012 seemed to suggest that a new GPU with a staggering "7.1 billion transistors" was on the way.
The first signs of a Kepler-based 7-billion transistor GPU were spotted in March 2012 - reported exclusively by HardwareZone.
In May that year, at GTC 2012, NVIDIA did in fact unveil the rumored GK110 GPU. But while details were sparse at the time, one thing was clear: GK110 was destined for the Tesla K20, an enterprise-class card costing in the range of US$3000. And with that, the dream of a GK110 GeForce card seemed over.
Until now.
Meet the NVIDIA GeForce GTX Titan
By far, the most powerful consumer graphics card ever made.
The GTX Titan takes its name from the Oak Ridge National Laboratory TITAN supercomputer located in Tennessee, USA, which in November 2012 was named the world's fastest supercomputer. Like the 18,668 Tesla K20X GPU accelerators found at the heart of the TITAN supercomputer, each GTX Titan is powered by NVIDIA's GK110 GPU, which comprises of 2688 CUDA Cores and 7.1 billion transistors, making it, by far, the most powerful GPU ever created.
To put that in perspective, that's 75% more CUDA Cores than the the GTX 680, and twice the number of transistors. In fact, the GTX Titan has slightly more transistors than even the dual-GPU GeForce GTX 690!
Core clock speeds on the Titan are set to 836MHz with boost core clock speeds reaching 876MHz using NVIDIA's new GPU Boost 2.0 technology (more on that below).
For those that complained about the GTX 680's lack of memory, the Titan also boasts 50% more memory bandwidth than the GTX 680 with a 384-bit memory interface, and a whopping 6GB VRAM, clocked at 6008MHz DDR (GDDR5). The massive graphics memory would come in handy in the most intense of situations such as surround gaming with 3D enabled on the highest possible game quality settings and when operating in SLI.
The reference card itself measures 266mm in length, which puts it right between the shorter 255mm GTX 680 and the longer 280mm GTX 690. Like the GTX 680, it will feature two DVI ports, one HDMI port and one DisplayPort output.
The GTX Titan will utilize the same ports as the GTX 680.
The GTX Titan will retail in the US for US$999, which is the same launch price as the dual-GPU GTX 690. NVIDIA says that both will share the top spot in its lineup. Local pricing has yet to be released, however, locally the GTX 690 ranges from about S$1480 to S$1580 so you can expect the Titan to cost about the same. With that kind of pricing and the expected performance throughput that should rival the GTX 690, but with new features that we would be highlighting below, we doubt the GTX 690 would still appeal to enthusiasts. We will of course ascertain its true caliber in our performance review coming up.
Here's a look at how the Titan stacks up against its competition:
[hwzcompare]
[products=373666,269094,298166,281584]
[width=175]
[caption=NVIDIA GeForce GTX Titan and competitive SKUs compared]
[showprices=0]
[/hwzcompare]
Under the Hood
For the first time, lower operating temperatures may actually translate to better in-game performance!
Reference cards aren't generally known for high-performance coolers or low acoustic levels, however, due to the complexity of the GK110 GPU, NVIDIA will not be allowing its add-in partners to customize or alter the reference card in any way, much like it did with the dual-GPU NVIDIA GeForce GTX 690.
Fortunately, NVIDIA has spared no expense in the thermal design of the Titan which primarily uses a high performance copper vapor chamber to provide its cooling. The vapor chamber draws heat away from the GPU using an evaporation process similar to how a heatpipe performs, but according to NVIDIA, in a more effective manner.
The heat from the vapor chamber is then dissipated by a large, dual-slot aluminum heatsink and an extended fin stack which extends beyond the base of the vapor chamber to increase the total cooling area. NVIDIA is also using a new thermal interface material from a company called Shin-Etsu that supposedly boasts over two times the performance of the thermal grease used on the GTX 680. Finally, the Titan also has an aluminum back plate to provide additional cooling for the PCB and a single fan to exhaust hot air.
Like the GTX 680, acoustic dampening material is used in the fan to minimize noise.
All of this is important because the GTX Titan will be the first card to use NVIDIA's new and improved GPU Boost 2.0 technology which, instead of using existing power overhead to dynamically adjust clock speeds like the original GPU Boost, now uses thermal thresholds.
GPU Boost 2.0
Originally launched last year with the GeForce GTX 680, NVIDIA's GPU Boost technology was designed to dynamically adjust core clock speeds by using excess available power budget to boost performance. The Titan's GPU Boost 2.0 works in a similar fashion, but uses a GPU temperature threshold instead. It will still consider the power drawn by the graphics card, but it becomes a secondary checkpoint and is no longer the main factor.
Similar to the original GPU Boost, GPU Boost 2.0 will automatically boost the Titan's core clock speed as long as it remains below a certain temperature - by default it is set to 80 degrees Celsius (but it can be raised higher yet). The Titan's GPU will constantly monitor temperature, adjusting core clock speed and voltage on-the-fly to maintain this target temperature.
The idea for this change is backed by the fact that the GPU can in fact operate safely at higher power draw as long as it's kept within its temperature threshold. Of course, there are upper limits to both and as long as they don't exceed the safe range, consumers stand to gain more.
Unlike the original GPU Boost, NVIDIA will now allow consumers to tweak GPU Boost behavior. As such, gamers that are comfortable with higher temperatures will be able to raise the GPU temperature threshold for higher clock speeds. Of course, higher temperatures also mean increased fan acoustics. Since fan noise is a big deal for many enthusiasts, NVIDIA's upcoming drivers for the Titan will also also allow you to control your preferred temperature target by meddling with fan noise levels (which govern the operating temperature, which in-turn limits your performance boost profile).
EVGA's Precision X showing the new GPU Boost 2.0 controls, with thresholds for temperature and power.
Furthermore, due to a change in ideology to how GPU Boost works and who the Titan is designed for, the power target settings in the control panel no longer defaults to the typical board power (which was 170W for the GTX 680), but it now sets the graphics card's maximum power draw (which is 250W at the 100% setting).
Hardcore liquid cooling enthusiasts should be excited by this news, as GPU Boost 2.0 looks like it will have great synergy with the low operating temperatures of a liquid cooled graphics card. That of course means getting your hands dirty to customize your own water block for this GPU, but the benefit for some enthusiasts can be immeasurable.
One more nugget of information is that GPU Boost 2.0 now comes with over-voltage control, allowing you to push forth for higher boost clocks bv means of increasing GPU voltage. The latter is in direct relation with the GPU's operating temperature, which is now in your direct control. Of course, to prevent permanent damage to the GPU, a voltage tweaking is limited to a safe range. However, over-voltage support by card vendors is optional and they can disable it in the video BIOS. As such, it remains to be seen which vendors will support this function.
For now, GPU Boost 2.0 will be exclusive to the GTX Titan.
Display Overclocking
VSync. Gamers have a long held love-hate relationship with it. On the one hand, it prevents graphics tearing sometimes seen when pushing high frame rates to a 60Hz refresh rate monitor, but on the other hand, it limits your game to a maximum of 60 FPS. GPU Boost 2.0 might solve that problem with a new feature called Display Overclocking. Essentially, this works by adjusting (overclocking) the pixel clock of your display, allowing you to hit higher refresh rates and thus higher FPS with VSync still enabled. Not all monitors currently support this feature and we haven't had the chance to test it ourselves yet, however, the idea is certainly intriguing. NVIDIA mentioned that it's basically a trial and error process, but if your monitor doesn't support it, you will see it default back to its original functionality.
Unlike GPU Boost 2.0, NVIDIA tells us that Display Overclocking will be available for older NVIDIA cards as well, however no time frame for the updated drivers was given.
Test Setup
For those who've been following our reviews, take note that we've replaced our old test rig with a brand new one for 2013. Here are the new specs we'll be running with:
- Intel Core i7-3960X (3.3GHz)
- ASUS P9X79 Pro (Intel X79 chipset) Motherboard
- 4 x 2GB DDR3-1600 G.Skill Ripjaws Memory
- Seagate 7200.10 200GB SATA hard drive (OS)
- Western Digital Caviar Black 7200 RPM 1TB SATA hard drive (Benchmarks + Games)
- Windows 7 Ultimate 64-bit
Here's the list of cards we'll be testing.
- NVIDIA GeForce GTX Titan 6GB GDDR5 (ForceWare 314.09)
- NVIDIA GeForce GTX 680 2GB GDDR5 (ForceWare 310.90)
- NVIDIA GeForce GTX 690 4GB GDDR5 (ForceWare 310.90)
- AMD Radeon HD 7970 GHz Edition 3GB GDDR5 (AMD Catalyst 12.11 Beta) - OC BIOS
- AMD Radeon HD 7970 GHz Edition 6GB GDDR5 CrossFireX (2-way) (AMD Catalyst 13.1) - OC BIOS
Note 1: There's a slight discrepancy in driver versions used across all the cards, but bear with us updating of benchmarks is an ongoing process, especially on our new test rig. Despite this, the differences shoudn't be alarming as they are fairly recent releases and still paint a reasonably accurate expectation of where the new card stands among the competition.
Note 2: In our performance testing, while the GTX Titan's GPU Boost 2.0 did work on the default settings, we were unable to produce higher scores by raising the target temperature using EVGA's Precision X tool. Despite setting Temp Target to 95 degrees Celsius (the max allowed), temperature still climbed no higher than 80 (the default setting). It is possible that this is a problem with the current ForceWare drivers or EVGA Precision X tool version we are using and a future update may be required. As such, all testing on the GTX Titan was performed with GPU Boost 2.0 left at the default 80 degrees Celsius threshold. We also understand that there's a secondary check in place to factor in the power draw, but given how we've tested and what we've observed, we tend to believe it's a limitation we've faced with regards to the early drivers or the utility.
Benchmarks
We've also updated our benchmark list, adding Unigine's new Valley benchmark, as well as 3DMark 2013 and the big one, Crysis 3! We've also updated Unigine from version 3.0 to 4.0.
- Futuremark 3DMark 11
- Futuremark 3DMark 2013
- Unigine 4.0 "Heaven"
- Unigine "Valley" 1.0
- Battlefield: Bad Company 2
- Crysis Warhead
- Crysis 2
- Dirt 3
- Assassin's Creed 3
- Call of Duty: Black Ops 2
- Hitman: Absolution
- Far Cry 3
- Crysis 3
3DMark 11 Results
As always, we start with 3DMark 11, our trusty synthetic benchmark designed to test a GPU's performance at various aspects of DirectX 11 such as tessellation and DirectCompute.
The Titan wasted no time in showing its prowess, outperforming the GTX 680 by 37% on Performance and an incredible 50% on Extreme. We were impressed by just how close the Titan is to dual-GPU level performance, as it trailed the GTX 690 by just 16% at Performance and 19% at Extreme.
Taking a close look at the GPU Boost 2.0 chart, you can see that the Titan's temperature rose to just under 80 degrees Celsius and is accompanied by a similar rise in GPU clock speed during the time frame when it was loaded.
GPU Boost 2.0 performance chart for GTX Titan on 3DMark 11 Extreme
3DMark (2013) Results
3DMark (2013)'s Fire Strike benchmark is Futuremark's most ambitious and technical benchmark to date utilizing a grueling two-part test with extreme levels of tessellation and volumetric illumination, as well as complex smoke simulation using compute shaders and dynamic particle illumination, while Fire Strike Extreme ramps up the difficulty with more tessellation, more particle effects and more taxing DirectCompute calculations.
We're still encountering some problems with dual-GPU setups on Fire Strike so for now we'll just be looking at single GPU performance. As you can see the Titan's GK110 GPU was no match for either the GTX 680 or HD 7970 GHz Edition, outscoring the HD 7970 by a massive 28% at Fire Strike and 35% at Extreme. This lead increases to well over 50% when you compare it against NVIDIA's previous single GPU flagship, the GTX 680.
Unigine 4.0 "Heaven" Results
Unigine's "Heaven" benchmark tests GPUs with extreme tessellation, dynamic sky with volumetric clouds, real-time global illumination, and screen-space ambient occlusion among others. This time round, we've updated the benchmark to the latest edition, version 4.0.
The Titan was again impressive, outperforming our single GPU cards by as much as 65% while trailing the dual-GPU setups by just 15-20% on average.
Unigine "Valley" 1.0
Valley is a new benchmark from Unigine that renders a highly detailed seamless landscape covering a total area of 64,000 square kilometers. It also utilizes dynamic sky effects, volumetric clouds, sun shafts, DOF, and ambient occlusion.
We saw a similar result to "Heaven" with the Titan performing about 60% better than the GTX 680 and HD 7970 GHz Edition while trailing the GTX 690 and HD 7970 GHz Edition CrossFireX setup by about 15-20%.
Crysis Warhead
For our first gaming benchmark we'll be using DX10 game Crysis Warhead. While it's getting a bit old, it does feature a fairly heavy amount of tessellation, making it a good starting point for testing the Titan.
The Titan showed it performs just as well for actual games, again outperforming the GTX 680 by about 55-60% while trailing the GTX 690 by just 15%.
Battlefield: Bad Company 2 Results
While Battlefield: Bad Company 2's DX11 Frostbite 1.5 engine is usually a fair test of a GPU’s geometry shading power, it was just no match for the powerhouses we're testing today with the Titan performing just as well as both dual-GPU setups. In fact, the performance is plateauing with the more powerful cards as it likely needs a well overclocked CPU to elevate the platform's performance. But with such high performance numbers to begin with, it doesn't really matter for single monitor users as any of these cards will give indistinguishable game experience - unless otherwise you can overclock your monitor's refresh rate to such high figures with Display Overclocking, but we highly doubt it. Let's move on to other gaming benchmarks.
Crysis 2
Moving on to our first truly challenging benchmark, Crysis 2's Ultra setting puts GPUs through their paces with a grueling test of tessellation, Parallax Occlusion Mapping, water rendering, and particle motion blur and shadowing.
The GTX Titan was extremely impressive at the 1920 x 1200 resolution, outscoring the GTX 680 by a huge 73% while trailing the GTX 690 by just 8%. At a resolution of 2560 x 1600 pixels, it still performed well, continuing to outscore the GTX 680 by more than 70%, but fell about 18% behind the GTX 690 and 37% behind the HD 7970 GHz Edition CrossFire setup.
In terms of actual gaming experience, there was noticeably smoother gameplay on the Titan compared to either of the other single GPU cards, whereas the difference between the Titan and the dual-GPU setups was very minimal. Given the kind of raw frame rate improvements, this was naturally expected.
Dirt 3 Results
We saw similar results in Dirt 3 with the Titan outperforming the single-GPU cards and trailing slightly behind the dual-GPU setups, but again with virtually no distinguishable difference between them in terms of actual user experience.
Assassin's Creed 3 Results
Surprisingly, the GTX Titan came out tops in Assassin's Creed 3, however we suspect that this is largely due to optimization issues with dual-GPU cards running this game. Even with newer drivers, we suspect some of those optimization issues may still linger as the GTX Titan underperformed compared to previous games, scoring only about 30% better than the GTX 680.
Call of Duty: Black Ops 2 Results
Black Ops 2 saw the GTX Titan return to form, outperforming the GTX 680 once again by 60% while trailing the GTX 690 by about 20%. As was the case with previous games where the GTX Titan scored over 100FPS, actual in game experience was indistinguishable between it and the two dual-GPU setups.
Far Cry 3
Far Cry 3 is running on the Dunia 2 Engine, a modified version of Crytek's CryEngine. As such, this benchmark features heavy tessellation, high amounts of volumetric lighting, and deferred radiance transfer volumes (global illumination).
This was the first game where we saw the Titan struggle and it failed to hit the target 60 FPS average on our tested settings, even dropping to below 30FPS at the resolution of 2560 x 1600 pixels with 8x MSAA. Far Cry 3 is however a very taxing game, so despite the Titan's low scores, it was still fairly impressive, outscoring the GTX 680 by as much as 75% while trailing the GTX 690 by only 10% on average. Turning off AA altogether would significantly increase performance scores.
Hitman: Absolution Results
Hitman: Absolution is another challenging benchmark, running on IO Interactive's proprietary Glacier2 engine, a very hardware intensive engine which is able to render up to 1200 NPCs at the same time. The benchmark tool used to test this game is extremely intensive and features a massive crowd of NPCs, Reflective Shadow Mapping (RSM), Direct Compute accelerated Bokeh Depth of Field, extreme tessellation and Ambient Occlusion.
The GTX Titan showed a similar test result to Far Cry 3, outperforming the GTX 680 by about 55% to 70% while trailing the GTX 690 by about 10%. At the 2560 x 1600 resolution, the game ran much smoother than on both the GTX 680 and HD 7970 GHz Edition, however, there were still some stuttering frame rates evident as the card wasn't anywhere as fast as the preferred 60FPS in our maximum test setting.
Do note that the benchmark tool used for this game is far more graphically intensive than the actual story mode campaign level it is based on.
Crysis 3 Results
Our newest gaming benchmark is the much anticipated sequel to Crysis 2 and, much like its predecessors, continues to be an absolute beast of a game, utilizing CryEngine 3 to the max with extreme amounts of tessellation, per-pixel per-object motion blur, Bokeh Depth of Field, displacement mapping on small terrain, particle and volumetric lighting and fog shadows, improved dynamic cloth and vegetation, dynamic caustics and diffuse shadows.
The GTX Titan may have finally met its match with Crysis 3 as it posted quite erratic scores, outscoring the GTX 680 by anywhere between 13% to 50%. It also performed quite poorly compared to the GTX 690 with AA turned off, dropping behind by as much as 45%. Surprisingly it did much better with 8xAA enabled, trailing by just 11%.
On a side note, both AMD cards performed quite poorly at Crysis 3, however a new driver set, Catalyst 13.2 Beta, was released this week which may address this issue. We'll revisit the AMD camp's performance with newer drivers in the next round of tests.
Overclocking Results
The GTX Titan was quite easy to overclock and for our efforts we were able to raise core clock speed to 1015MHz and memory to 6800MHz DDR. This resulted in a hefty 15% performance gain on Performance and 13% at Extreme on 3DMark 11, which put the GTX Titan very close to the performance of a reference GTX 690.
Temperature Results
As the GTX Titan is using NVIDIA's new GPU Boost 2.0 technology, temperature rose fairly quickly in each test, but was accompanied by a similar increase in core clock speeds. After hitting the threshold of 80 degrees Celsius, the temperature seemed to hold steady without any noticeable fan speed increases. The fan itself operates quietly until set to about 70% of its maximum operational level where it gets noticeably louder.
On another note, if GPU Boost 2.0 is active with all new NVIDIA graphics cards moving forward, temperature testing could become a moot point since the GPU would be eager to overclock itself to its best levels and maintain the default temperature threshold of 80 degrees Celsius.
Power Consumption Results
The GTX Titan showed quite high power consumption, on par with AMD's Radeon HD 7970 GHz Edition. In fact, at load it drew only 7% less power than the dual-GPU GTX 690 which was quite alarming for a single GPU card. While it may seem unsettling when compared against the NVIDIA camp of cards, it's still in positive light when compared to the AMD camp of cards.
One thing to note is that the Titan is likely operating at its best potential turbo clock speeds under the 80 degrees threshold, which means it will likely consume more power than usual. This is especially true when you consider that other than the Titan, none in the comparison operates in the same manner, thus making it seem more power hungry than expected when it was actually designed to push the boundaries.
Conclusion
The NVIDIA GeForce GTX Titan is one of the most impressive graphics cards we've ever seen. Its GK110 GPU is far and away the most powerful GPU ever produced, completely demolishing previous flagship GPUs from both NVIDIA and AMD. From a single GPU card perspective, there is nothing that even comes close to touching it in terms of performance. In fact, it's almost as good as the dual-GPU GTX 690.
The problem is, while its performance is almost as good as a GTX 690, its price is exactly the same as a GTX 690. So why would anyone pay the same amount for something that's almost as good as a GTX 690, when you can just buy a GTX 690 for the same price?
85% of the performance of a dual-GPU 690 at 100% of the price.
Furthermore, it's not as if the GTX Titan holds any other significant advantages over the GTX 690 either: it's not much smaller - any case that it will fit into, the GTX 690 should also be able to fit into - and its thermal operation is about the same too. As we've seen, power consumption is also only about 7% less. Based on its performance output, the Titan should be priced about 15% cheaper than the GTX 690 (roughly US$850 or S$1275). So why isn't it?
You Asked For It
If we had to speculate, it's because the Titan only exists because people have wanted it for nearly a year. Essentially, NVIDIA created it as fan service.
Consider that ever since NVIDIA released the GK104-equipped GTX 680 people have speculated over the possibility of a GeForce card with a GK110 GPU. The reason it hasn't been made before is because, realistically, it's neither cost-effective nor sound business sense for NVIDIA to do so.
The GK110 GPU is definitely not cheap to produce, which is why, up until now, it's only been available in the US$3000 Tesla K20 enterprise-class card (which itself has only been available since a few months ago). If NVIDIA releases the GTX Titan priced at US$850, not only does it lower its margins (in fact, based on the price of the Tesla K20, it might be making a loss), it also cannibalizes the sales of its own GTX 690 as enthusiasts looking for dual-GPU level performance may opt for for the slightly cheaper Titan instead.
But like we said, there is demand for the GTX Titan. So NVIDIA made it - but to do so, they've had to price it at US$999, giving you the option of either the GTX 690 or the GTX Titan. So while it doesn't fill any performance or price gaps, the GTX Titan does answer the question of "What if a GK110 GPU GeForce card existed?" Not to mention, it lets NVIDIA show off its engineering prowess.
Everything considered, if you're simply looking for the best performing NVIDIA card out there, for the same price, the GeForce GTX 690 is obviously a much better buy. But if the idea and prestige of owning a GK110-equipped card, the same GPU found inside the world's fastest supercomputer, appeals to you, you might be willing to take the slight performance hit to own a GTX Titan.
Our articles may contain affiliate links. If you buy through these links, we may earn a small commission.