The NVIDIA GTX 680 comes at a time when NVIDIA has found themselves constantly battling against AMD for the title of fastest graphics card. So far, AMD has done a pretty good job of keeping NVIDIA on their toes and delivering pretty good performance to price ratios. Admittedly, as a result of this war and the ever increasing size of chips and memory, prices of high-end GPUs have really only been inching upwards with some GPUs selling for about $600 a piece. We believe that this is a result of a few factors, primarily due to a lack of real capacity to manufacture these high-end chips as well as a low propensity to price them cheaply. As such, we’ve been seeing GPU price slowly inch upwards…

With the NVIDIA GTX 680, we are hoping to see a reversal of this trend. NVIDIA has completely changed the way that the company designs their chips in a whole host of ways and has built-in new features into these chips that make them a complete sea change from what NVIDIA has done in the past. NVIDIA, Since G80 (8800GTX) has really been all about big chips and that, for the most part, has really hurt them when it came to making their cards efficient, cool and inexpensive. AMD, on the other hand, has for the most part been all about smaller more efficient chips. Today marks the day when the tables have turned.

NVIDIA’s GTX 680 is based upon NVIDIA’s GK104 GPU which is part of the Kepler generation of graphics processors. The GTX 680 is a 3.54 billion transistor chip with 1536 shader cores which fits inside of a 294mm ^2 package which is manufactured on TSMC’s 28nm process. The card’s base clock is 1006MHz, and features a 1058 MHz ‘Boost Clock’ which means that the card can and will overclock itself with the Boost Clock when there is a light load on the chip. NVIDIA has opted for 2GB of GDDR5 running at 6Gbps or 1500Mhz clock before the quad-pumped multiplier effect of GDDR5. We will go further into the specifications of the GPU’s architecture and features in the following pages… All of this, fits into a chip with a TDP of 195w. Why does any of this matter? Because this is a significant departure from the GTX 580 (GF110) which is a 520mm^2 chip with 3 billion transistors (500M fewer) and 512 shader cores with a clock speed of 772MHz and a TDP of 244w. The GTX 680 (GK104) is a 43 percent smaller chip than the GTX 580 (GF110) and features 17 percent more transistors while tripling the amount of shader cores and reducing the power consumption by 20 percent.

In today’s review, we’ll be going a little over the architecture as well as the new features that the Kepler architecture and GTX 680 bring to gamers. We will also evaluate the GTX 680′s GPGPU performance as well as gaming performance, since after all, this is a gaming graphics card first and foremost. This review will be less technically focused on the GPU design decisions that NVIDIA has made and more of a focus on simple performance metrics and user experiences. We have written a separate piece solely dedicated to going in-depth into the architecture, the decisions made when designing that architecture as well as the future of the Kepler architecture as we see it.
The Architecture – Abridged

We are going to right off and say it quickly, this point of this review is not to banter on about the fine points of the architecture. If you’re interested in that, we recommend heading over to our article solely designated for a deep dive into the architecture and future architectures.


So, what exactly makes the Kepler Architecture different from, say Fermi? First, it has to do with the SMX unit which is NVIDIA’s update to their previous generation SM unit. The Kepler SMX features significantly more texture, geometry and shader processing capability than Fermi as each SMX has 192 shader cores per SMX unit compared to Fermi’s 32 shaders per SM unit. This results in 6 times as many shaders per SMX unit over Fermi with Kepler, and provides 2 times the performance per watt. Because of this, there are only 8 SMX units in the Kepler GK104′s architecture compared to 16 SM units in Fermi. So, what NVIDIA has done is increase the amount of shaders per unit by a factor of 6 but reduce the amount of shader units by half, netting the three times increase in shader cores from Fermi to Kepler.

In addition to the entire overhaul of the shader units, NVIDIA has also changed up their memory controller from a 384-bit memory bus in Fermi to a 256-bit memory bus in Kepler, which may be one of the reasons why NVIDIA was able to pack so many shader cores into such a small chip. They have also increased their memory speeds up to 6Gbps, which is one of the highest speeds in the industry for a graphics chip. Fermi operated on 4Gbps, which represents a 50% increase in speed, but due to the narrower memory bus, the memory bandwidth is actually slightly lower than Fermi (192.26GB/s vs 192.4 GB/s).

The Kepler architecture also features twice as many texture units, resulting in 128 texture units compared to Fermi’s 64 as well as more than double the texel fill-rate increasing from 49.4 gigatexels/s to 128.8 gigatexels/s. This increase is where NVIDIA’s Kepler architecture stands to significantly improve its graphical performance. It also supports Direct X 11.1 which NVIDIA doesn’t consider that big of a deal, but does put them on par with AMD.

The one thing we did notice was that NVIDIA had left out almost anything about Kepler’s HPC performance, which we found quite interesting considering the fact that Fermi was a huge improvement in terms of HPC performance in both single and double precision. NVIDIA only mentioned that Kepler’s single precision performance had doubled from 1581 GFLOPS to 3090 GFLOPS, but no mention of double precision performance… and as we will explain in our in-depth architecture analysis, there’s a reason why.
Kepler’s New Features

Card Physical Specifications
The GTX 680 has to be one of the smallest high-end GPUs that we’ve seen in quite some time. The GTX 680 is about the same size as an AMD Radeon HD 7870 and significantly shorter than the HD 7970 and GTX 590.
The GTX 680 dual-slot reference board measures 10" in length and includes two dual-link DVIs, one standard size HDMI port and one standard size DisplayPort 1.2. The card is powered by two 6-pin power connectors which can draw up to 225w when you add the two 6-pin connectors and the power coming from the motherboard. This card is only designed with a TDP of only 192w. The card also supports PCI-Express 3.0

Top to Bottom: HD 7970, GTX 590, GTX 680, HD 7870

NVIDIA claims that the GTX 680 only requires a 550 Watt PSU, which is 50w less than what they recommended for the GTX 580 even though the GTX 680 is poised to be much faster. The card also has a thermal threshold of 98C, even though we don’t anticipate it getting that high. We will detail exactly how much the card draws under idle and towards the end of our review.

GPU Boost
GPU Boost is a feature new to NVIDIA’s GPUs. GPU Boost is NVIDIA’s version of what has existed in the technology industry for quite some time, but never really in GPUs. Intel has their Turbo boost and AMD has their Turbo Cores. These boosting techniques purely exist when the processor isn’t under full load and can afford to clock up some of the cores of the processor beyond their base clock speed. These clocks are checked and modified every 100ms.

Similarly, NVIDIA has applied this principal to their GPU boost which uses available power on the GPU that isn’t being used to increase the clock speed. Since not all of the GPU is being used, there is additional thermal and power capacity on the chip and it dials itself upward in a variable manner in order to deliver better performance in applications that don’t necessarily fully tax the GPU.

So, the GTX 680 has a whole host of variable clocks that it runs at in order to reduce heat and power consumption. The base clock itself is 1006 MHz with a low power clock of 324 when the GPU is at 10 percent power and in an idle state. As the GPU begins to ramp, it will then hit the base clock of 1006 MHz and if the application so demands it, 1058, the boost clock. We have observed the card hitting 1106 MHz in some cases where the application allows it, as the GPU does not necessarily always limit itself to 1058 MHz as many have been led to believe. One must understand that when referring to the Boost Clock, NVIDIA is referring to the average clock frequency of the GPU as it will vary from millisecond to millisecond.

With NVIDIA’s Kepler, NVIDIA has been able to successfully enable FXAA in a single GTX 680 (GK104) whereas it used to take three GTX 580s to run the same settings with MSAA enabled. This is, of course, referring to the story we recently reported on during GDC 2012 where Epic Games reported to be running their Samaritan Demo on only one Kepler GPU vs three Fermis. FXAA isn’t necessarily a new technology as it already exists in some games, but it is being used extremely effectively on a single card with Kepler.

TXAA is NVIDIA’s new "film style" AA technique which is designed to exploit the GTX 680s high FP16 texture performance. TXAA is a combination of CG film style AA resolve and hardware AA wound into one. In the case of 2x TXAA there’s an additional optional temporal component for better image quality. There will be two modes of TXAA, TXAA 1 and TXAA 2. TXAA 1 will offer visual quality on par with 8x MSAA with a supposed performance hit of 2x MSAA, while TXAA 2 will deliver image quality beyond 8x MSAA but with the performance speeds of 4x MSAA.
Currently, no titles use TXAA technology and as was implemented with FXAA will be showing up in games later this year. Companies like Crytek and Epic Games will be supporting TXAA in their engines and games like Secret World, Borderlands 2 and Eve Online will utilize these technologies.

Adaptive Vsync
Adaptive Vsync is NVIDIA’s feature which enables the GTX 680 and Kepler architecture to reduce Vertical Sync stutters while also getting rid of screen tearing. Currently, the problem with using Vertical Sync or Vsync is that it either plays at 30FPS or 60FPS as those are the predetermined FPS caps designed to reduce screen tearing. If you remove Vsync, you can run at much higher frame rates, but you run the risk of having screen tearing.

The stuttering within using Vsync is a result of the GPU having to switch between running at 60 FPS down to 30 FPS and whenever it does this, it results in a stutter. NVIDIA’s Adaptive Vsync simply avoids this by smoothing out the frame rate drop and actually preventing the card from suddenly dropping down to 30 FPS. What it does is dynamically vary Vsync on and off so that when the frame rate is below 60 FPS it turns it off, and when the frame rate is above 60 it turns it on, effectively limiting it to 60 FPS and preventing tearing, but also preventing stuttering.

Kepler’s New Display Capabilities
With the GTX 680 and Kepler, NVIDIA has introduced some new much needed graphical capabilities. First off, NVIDIA now enables up to four displays to be run off of one card. In the past, in order to run more than two displays, you had to have two videocards. In addition to that, you could only run 2 displays per card and those cards could not be in SLI mode or you could run two graphics cards (GTX 580s) and have up to three displays in 3D Vision Surround (3 Displays in 3D mode).

Now, you can run up to 4 displays off of a single card (something that AMD has been able to do since the HD 6000 series and has enabled up to 6). As NVIDIA has been demonstrating, they enable a 3D Vision Surround setup with an additional monitor set atop. Or, you could simply run all four monitors in a span, but that is not very FPS friendly.

Graphics Card Outputs, Left to right: HD 7970, GTX 590, GTX 680, HD 7870

In addition to the 4 display gaming, NVIDIA has also added support for next-generation 4K and 3GHz HDMI displays, we are really excited for this as it gives us an opportunity to test the 4K display capability of the GTX 680 and HD 7970 against each other and see which company really has the 4K market ready to be supported.

NVENC is NVIDIA’s new hardware-level H.264 video encoder. This encoder is different from previous generations because it actually operates on the hardware rather than what has been done in the past. In the past, the ‘hardware encoder’ was software encoding done on hardware by doing all of the encoding using the shader cores/CUDA cores. Essentially using the compute capability of the CUDA cores to do the encoding through software. NVIDIA claims that this results in an almost four times faster encoder than the previous generation while consuming much less power.

NVENC enables full HD encoding up to 8x faster than real-time and supports H.264 Base, Main and High Profile Level 4.1 encoding. It also support MVC for stereoscopic video, which is an extension of the H.264 codec which is used for Blu-Ray 3D. In addition to that, NVENC enables 4096×4096 encode, which means that it supports 4K encoding.

NVIDIA currently states that the NVENC technology is available through proprietary APIs, and provide an SDK for development using NVENC. Later this year NVIDIA will enable CUDA developers to be able to use the high performance NVENC video encoder. We will be testing this encoder later in our review using Cyberlink’s MediaEspresso.
System Setup
In our system we will be running three different GPUs. The NVIDIA GTX 680, the AMD HD 7970 and the NVIDIA GTX 590. We will be running the majority of our benchmarks with these three cards included occasionally omitting the GTX 590 and occasionally adding other cards when the benchmarks allow us to.

Our system setup is as follows:
CPU – Intel Core i7 3960X ES
Motherboard – Gigabyte GA-X79-UD7 (with PCI-Express 3.0)
RAM – 16GB Quad Channel Kingston HyperX 1600Mhz
Storage – SSD – Patriot Pyro 120GB – HDD – Velociraptor 600GB
Power Supply – Thermaltake Toughpower XT Gold 1475w
GPUs – NVIDIA GTX 680 – NVIDIA GTX 590 – AMD HD 7970

GTX 680 – 300.99
GTX 590 – 296.10
HD 7970 – 12.2

Benchmark Setup
We will be running the following benchmarks: SiSoft Sandra 2012, Futuremark 3DMark 11, Unigine Heaven 3.0 and Cyberlink MediaEspresso. We will also be benchmarking the following games: Battlefield 3, Batman Arkham City, Counter-Strike: Global Offensive, DiRT 3, Metro 2033 and Skyrim. In our testing, all of the games will be running at absolutely maximum settings while running at our monitor’s native resolution of 1920×1080. This means that we will be cranking up every single game to its maximum potential keeping in mind that things like CSAA and FXAA will not be part of our graphs as they are NVIDIA exclusive technologies and do not fairly compare against AMD’s running MSAA. As such, all of our testing will be with MSAA and set to the maximum settings that the game enables.
SiSoft Sandra 2012
In SiSoft Sandra 2012, we will be testing the GPU’s various compute capabilities, GPGPU, as well as its rendering and CUDA peformance.

First off, we’ve got GPGPU Processing running straight off of the compute shader. In SiSoft Sandra, everything is measured in Mpix/s which is how they benchmark compute capability. Here we’ll be comparing the GTX 680 vs the HD 7970. As you can see, in float shader (single precision) the GTX 680 delivers 3,380 MPix/s while the HD 7970 delivers 3,200 MPix/s so in this case the GTX 680 out performs the HD 7970 by a small margin. In Double Shader (double precision), though, the tables are turned significantly and the GTX 680 only reports 114.79 MPix/s while the HD 7970 delivers 662.32 MPix/s.

Moving onto OpenCL processing capability we once again see a similar trend. Except, in this benchmark, the AMD HD 7970 out performs the NVIDIA GTX 680 in both float shader and double shader performance tests by a pretty significant margin in both tests. In the Float test, the GTX 680 scores 3,006 MPix/s while the HD 7970 delivers.

In Compute Shader Cryptography, the gap widens further with the GTX 680 only scoring 6.548 GB/s in Encryption/Decryption Bandwidth, while the HD 7970 scores 29.869 GB/s. This trend continues in the Hashing Bandwidth benchmark with the GTX 680 delivering 9.93 GB/s while the HD 7970 delivers 23.729 GB/s. Looking at these results, the GTX 680 does not appear to be as useful of a compute GPU as we might have thought, which may have to do with NVIDIA’s future plans.

In OpenCL Cryptography, though, NVIDIA does once again show its single precision advantage by quite a bit by returning the favor to AMD. AMD beat NVIDIA by about 5X in Compute Shader Encryption/Decryption Bandwidth, and NVIDIA beats AMD by about 5X in OpenCL Encryption/Decryption Bandwidth. The GTX 680 delivers 10.889 GB/s in Encryption/Decryption Bandwidth while the HD 7970 offers 2.363 GB/s. The tables, though, do get flipped in OpenCL Hashing Bandwidth where the GTX 680 offers only 4.62 GB/s of Hashing Bandwidth, while the HD 7970 delivers 18.966 GB/s.

In Video Rendering, we see the GTX 680 deliver 1,069 MPix/s in Float Shaders while the HD 7970 beats it with 1,456 MPix/s. In the Double Shaders benchmark, we see a continuation of this trend, with the GTX 680 delivering 114.79 MPix/s and the HD 7970 quadrupling it with 471.35 MPix/s. Needless to say, the 7970 handily wins this benchmark.

Last but not least, we have CUDA processing capability where we compare the GTX 680 against the GTX 590. In this benchmark we see that the Float Shader test yields the GTX 680 with 2,702 MPix/s slightly edging out the dual Fermi GPU GTX 590 with 2,610 MPix’s indicating a huge performance increase over the Fermi generation of GPUs in terms of single precision performance. When it comes to the Double Shaders test, though, the GTX 590 beats the GTX 680 by putting up 471.35 MPix/s to the GTX 680′s 204 MPix/s.

Looking at our GPGPU performance and results against the HD 7970, the GTX 680 struggles significantly in double precision tests. The only benchmarks where the GTX 680 doesn’t get handily beaten in are in the tests where its single precision capabilities outpace those of the 7970.
Futuremark 3DMark 11

In Futuremark’s 3DMark 11, we tested all three different levels of the benchmark’s performance against all of the cards we had in our arsenal to see which cards performed the best. We will also be using 3DMark 11 to measure our overclock performance and scaling as well.

In 3DMark 11 we got some interesting results. The GTX 680 outperformed the GTX 590 (and every subsequent GPU) in the Entry level benchmarks with a score of E14517 while the GTX 590 scored E14512 and the HD 7970 E12185. In the Performance benchmark, though, the GTX 590 outperformed the GTX 680 by scoring P9851 while the GTX 680 scored P9738, a difference of about 1-2 percent. Once you take into consideration that you’re basically comparing two Fermi GPUs in SLI against a single Kepler GPU, you realize that this is actually extremely impressive that the GTX 680 and GTX 590 are trading punches in a synthetic benchmark. The HD 7970 comes in 3rd at P8083 in the Performance test. In our Extreme test, the most intensive of all the 3DMark 11 benchmarks, the GTX 590 edges out the GTX 680 once again by a few percentage points with the GTX 680 scoring X3177 and the GTX 590 X3307. Once again, the HD 7970 came up in 3rd place with X2756.

Looking at these results, we can definitely see the GTX 680 shining above all other competition with it effectively performing at the same levels as a GTX 590. Since the GTX 590 is a dual Fermi part, it is pretty nice to see NVIDIA’s single GPU part beating and keeping even with NVIDIA’s previous generation dual GPU part. This doesn’t even take into account that the GTX 680 consumes about 20% less power than the GTX 590 and HD7970.

Unigine Heaven 3.0

In Unigine Heaven, we decided to go all out and not run the benchmark at the same settings that everyone else runs their as at. Instead, we opted to run the benchmark at its absolutely full capacity. That means that we went and turned up every possible setting we could to its maximum setting. We ran the DX11 benchmark at 1920×1080 (not the standard 1600×900) and we turned Tessellation to Extreme, Shaders to High, Anisotropy to 16X, Stereo 3D Disabled and Anti-Aliasing to 8X all while running full screen.

During our benchmarks, we noticed that the GTX 590 actually delivered the best overall performance above both the GTX 680 as well as the HD 7970, but we noticed significant jitteriness in the frames as well as jitters in the shadow movement. Even though both the GTX 680 and HD 7970 performed lower than the GTX 590, the GTX 680 really had the smoothest experience. Looking at the graph above, we can see that the HD 7970 had a minimum FPS of 23.6 while the GTX 680 came in at 25.9 and the GTX 590 at 26.5. The average frame rates were 46.9 for the GTX 680and 51.3 for the GTX 590, with the HD 7970 coming in even at 43. Max FPS were 103.3 for the HD 7970 as well as 120.1 for the GTX 680 and 121.7 for the GTX 590.
Cyberlink MediaEspresso

The reason why we tested with Cyberlink MediaEspresso is because it features GPU accelerated encoding. This program uses NVIDIA’s new NVENC codec which enables us to most accurately test NVIDIA’s new video processing architecture. This program has both hardware and software acceleration, so what we’ve done is enable it and disable it for each GPU when paired with our Intel Core i7 3960X. This was by taking a 1080P video file and converting it to 720P with a very high bit-rate.

Looking at our benchmarks, you can see that the GTX 680 actually performs the best above all GPUs and even against an Intel Core i7 3690X. The GTX 680 did the conversion in 19 seconds while the CPU did it alone in 23. The GTX 590 did it in 36 seconds and the HD 7970 did it in 34 seconds. Interestingly, though, is that the non-hardware accelerated performance was faster with the 7970 than it was with the GTX 680 or GTX590. So, with this application at least, you either want to use the GTX 680 or just the CPU itself.
Gaming Benchmarks
In our game testing for this review we decided that since we were playing with the three fastest cards in the world that we would have to run all of the games at their absolute maximum settings and see how each card fared individually. Based upon our findings here, we would see which card was really the king of the high-end market. All of our benchmarks listed below were run at 1080P resolution and with the game’s settings set to the absolute maximum settings that made it comparable to other cards.

Battlefield 3
In Battlefield 3, our favorite game, we wanted to see how the GTX 680 faired against the HD 7970 and the GTX 590. Looking at our results, we can see that the GTX 680 not only wins in the average frame rate category, but also in the minimum frame rate category. To us, minimum and average frame rates are the most importance as they have the biggest influence on game performance.

The GTX 680 had an average of 94 FPS while the GTX 590 had an average FPS of 79 and the HD 7970 had 65 FPS. For minimum FPS, the GTX 590 actually did very poorly with 16FPS with the HD 7970 at 37 FPS which is still very playable. The GTX 680 is ridiculously higher than both of those cards with a minimum FPS of 87 FPS which is amazingly high considering the settings that we had been running it at. If you look at the graph itself, you’ll notice that the GTX 680′s Max, Min and Average FPS were all relatively close to each other, unlike the GTX 590. We also noticed some artifacting and various graphical glitches on the GTX 590 that we did not experience on the other cards.

Batman Arkham City
In Batman Arkham City, we tested this extremely heavy hitting game (no pun intended) with our top three cards, but we also decided to throw the GTX 680 at 32XCSAA which is an NVIDIA only setting for their graphics cards on the GTX 680.

Looking at our results, the GTX 590 performed the best in terms of average and minimum frame rate and the game played pretty smoothly. The GTX 680 performed at a minimum FPS of 29 while the GTX 590 played at 31 FPS and the HD 7970 played at 26 FPS. In our eyes, anything under 30 FPS will have noticeable lag, but when playing this game we didn’t really notice any on any of the cards even though the frame rates did dip below 30 or hover around 30. The average frame rates for the cards were 53 for the GTX 590 and 45 for the GTX 680 as well as 42 for the HD 7970, essentially saying that the 7970 and GTX 680 had a similar level of performance in this game. Admittedly, this is an NVIDIA favored title, so that should be factored into the equation as well.

We also tested the GTX at 32X CSAA and took a slight hit to FPS with the minimum FPS going down to 25 FPS and the average actually remaining flat at 45 FPS. This simply means that even at a higher level of AA, the GTX 680 handles Batman Arkham City pretty well for being a single GPU. But we are still a bit disappointed in the sub 30 frame rates. Hopefully with future driver updates we can see some performance improvements to improve the minimum frame rate of the GTX 680.

Counter-Strike: Global Offensive
Since Counter-Strike: Global Offensive is the latest iteration of Steam’s Source engine, we figured it would be the best way to test all of these high-end graphics cards even though we know that the source engine is not that graphically intensive.

In CS:GO the GTX 590 actually won the benchmarks for minimum and average frame rates. The GTX 590 pulled in an average of 258 FPS while the HD 7970 ran at 245 FPS and the GTX 680 ran at 170 FPS, significantly lower. We could chalk this up to drivers, but we shall see. As it stands right now, all three cards had at least 100 FPS and if you ask anyone that’s enough to play smoothly. Also, all three cards hit the steam engine max FPS of 300 as well.
DiRT 3
With DiRT 3 we figured that we would pick an AMD promoted title which also shares the same engine as the Formula 1 games which are also Codemaster’s games. As such, we knew what kind of frame rates we could expect, so we of course turned up all the settings to their absolute maximums once again.

When it came to average frame rates, the GTX 590 slightly edged out the GTX 680 with an average of 96 FPS while the GTX 680 ran at 92, a relatively negligible amount. The HD 7970 came trailing in 3rd place at 80 FPS, which we found extremely odd considering that this is an AMD promoted title and DiRT has consistently been a launch title for all of AMD’s GPU lines. So, when we found these frames we were very confused until we realized that for some reason, as AMD has been updating their drivers the performance of DiRT 3 has been dropping by a few frames per update until you realize that the difference can be over 10 frames… The minimum frame rates were very close to their average frame rates with the GTX 680 and GTX 590 actually being tied at 85 FPS and the HD 7970 coming in at 74.

Metro 2033
Traditionally, we have considered Metro 2033 to be the test that brings all videocards to their knees, especially at absolute maximum settings. With Metro 2033, we were confident that we would put quite a bit of stress on the GTX 680 as well as the GTX 590 and HD 7970. We once again, like in all of our other previous tests, set the game to absolute maximum settings and attempted to play the same level.

In our testing of Metro 2033, we found something extremely interesting. The HD 7970 actually had the best average frame rate as well as the highest minimum frame rate, which we found extremely interesting since Metro 2033 is a TWIMTBP title. Admittedly, the difference between the three cards was so narrow that the margin of error should be taken into account since the HD 7970 had an average frame rate of 55 FPS while the GTX 680 had 54 and the GTX 590 had an average frame rate of 53. So, a delta of 2 frames really doesn’t mean that there’s really any winner. We would’ve expected the GTX 680 to have performed faster than the HD 7970, but we could chalk that up to the fact that these are the first drivers for the GTX 680 and there could be significant performance improvements in the future.

Also, when looking at the minimum frame rates, we noticed that both the GTX 590 and the GTX 680 reported a minimum frame rate of 24 FPS while the HD 7970 reported a minimum of 28, admittedly all three cards are under our target of 30 FPS, but the HD 7970 and GTX 680 did not have any noticeable lag. The GTX 590, on the other hand, lagged like crazy and was, at parts, almost unplayable which we found extremely interesting especially after we looked at the data and noticed relatively close all three cards were to each other in terms of performance.
Frankly, we’d call Metro 2033 to be a tie, but nevertheless, a very weird result and we really expected more out of the GTX 680 in such an NVIDIA promoted title. Perhaps they put the majority of their driver optimization for this round into Battlefield 3 and other games and not something as old as Metro 2033? We just hope they improve the frame rates a bit, even though we really didn’t notice any lag.

If you’ve read our previous benchmarks, then you know what to expect from us in this benchmark. We once again ran Skyrim at 1920×1080 resolution with absolutely everything maxed out and made sure that we had a comparable setting across all three cards. We also tested FXAA on the GTX 680 just to see the performance hit.

In this test, the GTX 680 absolutely pulls away from the other two graphics cards by a quite noticeable margin. When it came to average frame rate, the GTX 680 came in at 84 FPS while the HD 7970 came in at 60 FPS and the GTX 590 at 59 FPS. Interestingly, when it came to minimum FPS, the cards were much closer with the GTX 680 running at 43 FPS and the HD 7970 at 43. Even though we don’t really consider maximum FPS a huge factor in gaming, the GTX 680 hit a max FPS of 155 while both the GTX 590 and HD 7970 ran at a max of 64. Mind you, we had to go into the game files and modify a text document in order to disable Active Sync in order to measure FPS beyond 60 since the game doesn’t actually come with an option to disable Active Sync.

We also tested Skyrim with FXAA on and the GTX 680 had a minimum frame rate of 39 (instead of 44) and an average frame rate of 76 (instead of 84). So, it does appear that FXAA when compared to standard 8x MSAA did put more of a toll on the card, but still played extremely well.
Overclocking and OC Software
For overclocking and temperature monitoring, NVIDIA supplied reviewers with EVGA’s latest version of their Precision overclocking utility. In this case, the newest version is called Precision X. Precision X is without a doubt, a huge departure from previous version of EVGA’s precision utility and it’s good to see that NVIDIA is recognizing EVGA’s awesome utility which has sparked many of their competitors to do the same thing, giving consumers added value for their purchase.

When it came to overclocking the GTX 680, we had a lot of fun figuring out how to properly overclock the GPU in a way that would actually result in better performance without killing the card. The good thing is that the GTX 680 has a lot of overclocking headroom, even on the reference cooler with the reference PCB design and VRM. Considering what we were able to do to this card with EVGA’s Precision X utility, there’s no doubt that some manufacturers with custom PCBs and proper VRMs could easily beat our overclocks.

In order to actually overclock the GPU, you have to increase the GPU clock offset, which increases the curve by which the base clock and boost clock operate. By doing this you are shifting the curve of the power target, base clock and boost clock. But of course, you can’t simply overclock the GPU frequency without increasing the power target, so we of course started to mess with that as well. After a lot of trial and error we were able to get a stable OC that was also benchable.

In our testing, we were able to get the card to run at 1,316 MHz and at that frequency we were able to benchmark 3DMark 11 in both Entry, Performance an Extreme levels. Considering the fact that the GTX 680′s stock clock is 1,006 MHz we were able to overclock the card by about 30% on a reference cooler and stock cooling. This is simply astonishing for a flagship card. We have a graph below that illustrates the overclocking increases and potential of this card.

Looking at the scores, you can see that the GTX 680 without a doubt takes the performance crown from the GTX 590 and when compared to the stock clocks of the GTX 680, yields considerable performance improvements. The Extreme score of E3732 is an increase of 17 percent and the Performance score of P11133 resulting in a change of 14 percent while our Entry score of E15783 only netted 9 percent since the Entry test is much less GPU bound than the Extreme test.

Power Consumption and Heat
In terms of heat and power consumption, the GTX 680 was absolutely stellar. At idle the card ran at only 324 MHz and as a result generally consumed around 10% of TDP or 20 watts. The card would generally idle around 31C and was so cool that no warm air actually came out of the back of the card. This was the most astonishing thing about this card, normally, under idle you get some sort of heat coming out of the back of the card but the GTX 680 has NONE.

Under sustained full graphical load, the card consumed at a maximum, 186w but generally hovered more around 172w. Under these scenarios, the card got as hot as 76C and never hotter. The fan speed of the GPU also stayed extremely quiet even under heavy loads, although, we did notice that different programs loaded the GPU differently and that the temperature of the GPU did not always dictate the fan speed.

We did, against the best wishes of NVIDIA, run Furmark and were only able to get the card to peak at 80C meaning that this card is inherently very cool and for you to break 80C you’ve either got to be doing something crazy with terrible ambient temperatures or you’ve got a bad card.

Out of all the three cards we tested, the GTX 680 was by far the quietest and coolest. It was followed by the HD 7970 which was also a pretty quiet card until it got put under a heavy load, then it became both loud and hot and felt like a blowdryer. The fastest the fan on the GTX 680 ever got was 54% speed, which was handily abused by us when we set the card to 75% when we were overclocking. The card was designed to be extremely cool and quiet and NVIDIA did a great job with that. We’re not even going to talk about how insanely loud the GTX 590 got under full load or how loud its idle was compared to the other two cards.
So, what kind of value does the GTX 680 deliver? Well, considering that it is supposed to be NVIDIA’s flagship card (which is debatable based on our architecture analysis), it is priced pretty high at $499. Many expected the card to be selling for $550 which is the same price as the HD 7970 launched at. What should be interesting to see is how AMD will adjust their pricing to reflect the GTX 680′s pricing and performance that we have seen here.

Newegg Jumping the Gun Letting us Show You All the Prices

Considering the performance that we got out of this card as well as the overclocking potential that we were able to unlock merely by using EVGA’s Precision X there is no doubt that this card is a great value, even at $499. Also taking into consideration that it replaces what normally would require a much more expensive GPU or two slightly less expensive previous generation GPUs, there is no doubt that this card is quite a value.

The GTX 680 came to us with extremely high expectations, especially after what NVIDIA had said about the HD 7000 series. There is no doubt that the GTX 680 is definitely the best gaming graphics card on the market today, and considering the amount of performance it delivers at the power levels and heat levels that it operates on it is without a doubt the best.

The GTX 680 handily beats its predecessor the GTX 590, a dual Fermi GPU card, and also beats its competition in most benchmarks (admittedly, not all). Not only does the GTX 680 do these things faster, cooler and quieter than all of its competition, but it does it for a lower price and in a smaller package (shorter board).

Looking at some of our benchmarks, it does appear that the GTX 680 does need some work on the driver side of things just to improve some performance here or there. But from what we’ve seen in our benchmarks and our gaming experience, there were no poor gaming experiences with the GTX 680. We just hope that NVIDIA can continue to further improve performance in games where the GTX 680 is slower than the 7970, namely games like Metro 2033.

Taking all of the factors we’ve talked about, performance, overclocking, heat, power consumption and design, there is no doubt in our minds that the GTX 680 wins our Editor’s Choice Award as well as our Innovation award. Yes, we’re giving it two awards.