Return of the King
With the GeForce GTX 780 Ti,
has snatched the single-GPU performance crown back from the clutches of the recently launched Radeon R9 290X, and not just by a small margin either, but by a landslide. By dethroning the R9 290X Nvidia has also taken the GTX Titan to the woodshed as well, as the GTX 780 Ti is far and away the fastest single GPU we have ever tested. Read on to see how it fares against the GTX 780, the R9 290X, and the former champ, the GTX Titan.
The Real Big Kepler
Back when the
GTX Titan launched
we all proclaimed it to be "
," or the full implementation of the Kepler architecture instead of the half-Kepler GK104 we got with the
. Of course, we all loved the GTX 680 at the time, but it was roughly half the size of the GK110 chip Nvidia had deployed to supercomputers worldwide. When Nvidia finally got around to stuffing the GK110 into a gaming GPU named Titan, we all rejoiced since we had finally acquired the real-deal Holyfield Big Kepler GPU.
It's hard to notice in this image, but the cooling shroud has a darker, smoked appearance to match the darker lettering.
However, even the Titan wasn't a full GK110 part, as it had one of its SMX units disabled. This begged the question - would Nvidia ever release a Titan Ultra with all SMX units intact? With the GTX 780 Ti we finally have that card. Not only does it have all 15 SMX units enabled, this bad mutha also has the fastest memory available on an Nvidia GPU with its 3GB of 7GHz GDDR5 RAM. Previously, this speed of memory was only found on the mid-range GTX 770. The bottom line is Nvidia is pulling out all the stops with the GTX 780 Ti in an effort to shame the
, and once again establish itself as the king of the single-GPU space. It should be noted that the GTX 780 Ti does not offer Double Precision compute performance like the GTX Titan, so CUDA developers will still prefer that card. The GTX 780 Ti is made for gamers, not scientists. We should also point out that the GTX 780 Ti supports quad-SLI, just like the GTX Titan, and the GTX 780 does not.
GTX 780 Ti Specs
Let's have a look at the specs of the GTX 780 Ti along with its closest competitors.
*The R9 290X's TDP isn't a quoted spec from AMD but rather one with air quotes around it. We believe it to be a bit higher than 250w.
On paper it's clear the GTX 780 Ti has a higher specification than either of its competitors, not to mention the obvious GTX 780. Although its memory bus isn't as wide as the R9 290X's, it has faster memory, so it's able to achieve higher overall memory bandwidth. The R9 290X is capable of pushing 320GB/s thanks to its slower 5GHz memory but wider 512-bit channel, while the GTX 780's faster 7GHz memory can squeeze 336GB/s through its narrower 384-bit bus. The GTX 780 Ti has more processing cores as well, and thanks to Kepler's higher level of efficiency compared to AMD's GCN architecture, is able to sustain much higher clock rates at all times as well. All that adds up to one ass-kicking GPU, as we'll see shortly. Like the GTX 780 the card measures 10.5 inches in length, and requires a six-pin and an eight-pin power connector. TDP is unchanged at 250w.
What's New Compared to the GTX 780
Since this board carries the GTX 780 moniker, let's look at how it is different from the GTX 780, because remember, this card costs $200 more than the original GTX 780 now that Nvidia has
lowered its price
. First up, it has 25 percent more CUDA cores, going from 2,304 to 2,880, which is quite a jump. Second, it has faster GDDR5 memory, which has been bumped up a full 1GHz to 7GHz. Third, it has a new feature Nvidia calls Max OC that simply balances the power going to the card from its three sources: the six-pin and eight-pin rails, and the PCI Express bus. Nvidia claims the board usually does this on its own quite well, but when overclocking all bets are off and not enough power from one source could limit the overclock. It claims this situation is rectified on the GTX 780 Ti, so you should be able to overclock this board higher than you could a GTX Titan or GTX 780. Finally, though it's not a new feature, this card also supports GPU Boost 2.0, like the other cards in the 700 series. However, with the arrival of the variable clock rate Radeon R9 290X, Nvidia is pointing out that it guarantees a base level of performance on all its 700 series cards, regardless of operating conditions. This is in contrast to the new Hawaii boards from AMD, which state a "max clock speed" but not what the actual average clock speed is under load as it tends to be a bit lower. We'll have more on that a bit later.
One of the most interesting features Nvidia has announced recently for its Kepler GPUs is
, which is technology built into upcoming LCDs that enable it to work hand-in-hand with the Kepler GPU to sync refresh rate and frames coming out of the GPU. It's essentially the end of V-sync as we know it, and since most hardcore gamers never use V-sync we couldn't be more thrilled about this technology. By syncing the monitor's refresh rate with the actual framerate coming out of the GPU, tearing and sheering is totally eliminated, resulting in a much smoother visual experience on-screen. There are some caveats, of course. First, we have not tested or witnessed G-Sync in action in our own lab, and have only seen an Nvidia-prepared demo of the tech, but what we've seen so far looks very good, and we have no reason to doubt it won't fulfill its promises once it lands in the lab.
In order to experience Nvidia's G-Sync technology you'll need a G-Sync LCD. The first one from Asus is a $400 24" model.
However, since we haven't seen it yet as the monitors are not yet available, we'll have to wait to deliver a verdict on this particular piece of gear. Second, in order to acquire this technology you will have to first acquire a G-Sync display, or buy an actual PCB and mod your monitor somehow. We're not sure how that would work, and what monitors will allow it, so again, we'll have to wait and see. We don't believe most gamers will want to buy a new LCD just to get this technology, however. Still, kudos to Nvidia for taking on a problem that has existed for as long as we can remember. If it really is as good as John Carmack and Tim Sweeney say it is, it could revolutionize the gaming industry. We'll have to wait and see.
ShadowPlay is more efficient than FRAPs, and doesn't consume your entire hard drive either.
We covered this technology at the GTX Titan launch, and back then it was "coming soon." Now that it's finally out, though still in beta, this is technology exclusive to Nvidia that should factor into one's purchasing decision. Since we've already covered it, in brief it lets you capture gaming footage with almost no performance penalty, according to Nvidia. Once captured the onboard H.264 encoder built into the Kepler architecture compresses it to reduce file size, and it works in the background always recording what you last did in the game, hence its name. We have been playing with it in the lab, so expect a writeup on our experience with it shortly.
Hit the second page for a discussion of heat, power, overclocking, benchmarks, and our final thoughts.
Heat, Power, and Overclocking
We'll cover the R9 290X "Golden Sample" controversy below, but for now let's focus on the GTX 780 Ti. Like all Kepler cards it runs very cool, and very quiet. Even with its extra cores and faster RAM it is typical to see it hit about 82C under load, and at that temperature it was barely audible in testing. This is the exact same experience we had with the GTX 780 before it, and the GTX Titan as well. These cards run very quiet, and never get too hot. And now that the R9 290X is out, the Nvidia cards seem downright chilly by comparison.
As far as overclocking is concerned, we've always had a very easy time overclocking Kepler boards, and the GTX 780 Ti was no different. Though Nvidia claims this board overclocks better than the GTX 780 and GTX Titan thanks to its load-balancing tech, we didn't experience that. Instead we achieved results which were just a tad bit lower than what we experienced with boards like the Asus GTX 780 DC2 and EVGA GTX 780 ACX. Overall we were able to hit 1,225MHz boost clock with a 250MHz memory overclock, which is pretty damn good. When overclocked the board hit 85C and had its fan spinning at 67 percent, though it was quieter than the R9 290X fan at 49 percent. Keep in mind we were unable to overclock the Radeon R9 290X since out of the box in its default "quiet" mode it hits 94C quite easily, leaving no headroom for overclocking. Sure, the R9 290X is already running at or around 1,000MHz during normal operation, which is higher than the stated Boost clock for the GTX 780 Ti, but in reality the R9 290X's typical clock speed is more around 950MHz or so. Nvidia would say it's actually around 800MHz, but more on that later.
Our default resolution for cards of this stature is 2560x1600 with 4XAA enabled, and all details fully maxed out. We play with everything turned up as high as possible, because, well, this is Maximum PC you are reading. Let's examine the numbers:
Best scores are bolded. Our test bed is a 3.33GHz Core i7 3960X Extreme Edition in an Asus Rampage IV Extreme motherboard with 16GB of DDR3/1600 and a Thermaltake ToughPower 1,050W PSU. The OS is 64-bit Windows 8. All games are run at max everything at 2560x1600 with 4X AA except for the 3DMark tests.
Now then, with the numbers in front of us we can begin to explore the complicated question of where these three cards stand in the current leader boards. We are just kidding, of course, because one look at this chart and one thing is immediately clear. The GTX 780 Ti kicks the crap out of everything, by a lot. We're used to seeing a few frames per second difference between one card and another when comparing cards of the same generation, but the GTX 780 Ti is just in a league all by itself. Nothing else even comes close, not even the mighty Titan, which costs $300 more. Of course, the R9 290X costs $150 less, so there's that to consider, but the end result from these tests is one simple statement -- Nvidia makes the fastest single GPU in the world, period. Unless AMD has a new piece of silicon that is even faster than Hawaii up its sleeve, which would be pretty amazing if it were true, it will be handing the fastest GPU crown back to Nvidia for the time being. We imagine Nvidia will hold onto this title for awhile now too, as AMD can't push the R9 290X any further than it already has. We suppose a water-cooled R9 290X or super-air-cooled version could boost performance a bit, but the best AMD could hope for would be to match Nvidia's card. We doubt it will be able to beat it any time soon.
With a card this powerful, you can certainly run most of the latest games at 4K resolution. And if you have the type of cash to spring for a $700 GPU, you might have the $5k or so required to land one of these sexy LCDs on your desk. Our hats are off to you, rich PC gamer, as gaming in 4K is truly breathtaking. Okay, here are the numbers:
Best scores are bolded. Our test bed is a 3.33GHz Core i7 3960X Extreme Edition in an Asus Rampage IV Extreme motherboard with 16GB of DDR3/1600 and a Thermaltake ToughPower 1,050W PSU. The OS is 64-bit Windows 8. All games are run at 3840x2160 with max everything, AA turned off. We do not have scores for the GTX 780 with Batman as we ran out of time to test it, but will update this chart ASAP.
At 4K the GTX 780 performs quite well but not as well as the more expensive Titan, and it also performed slightly worse in Battlefield 3 than the R9 290X. That said, the reviews of the R9 290X and the R9 290 generally showed the AMD cards performing better than their Nvidia counterparts at 4K. As we stated in our review of the R9 290X, AMD sent us a 4K panel in order to highlight this advantage it had over Nvidia, presumably due to their card having higher memory bandwidth and more memory too. However, with the GTX 780 Ti that advantage has largely been wiped out. However, it's worth keeping in mind that the $550 R9 290X performed quite well at 4K against its more expensive competition from Nvidia, so in a way it still holds a slight advantage, at least at this resolution. That's not worth very much in the real world though, as we can't imagine many people are gaming at 4K yet. It's just too expensive at this time, though it's amazing that a single GPU can run the latest games at decent frame rates at this resolution. We are truly living in an amazing time given all the GPU power at our disposal.
A Final Note on Heat, Noise, and Performance
A lot of ink has been spilled this week, at least digitally, on the heat, noise, and power consumption of the card that dethroned the GTX 780, the Radeon R9 290X. The reason for all the hub bub is two fold. First, AMD doesn't state a base clock for this GPU like it has done with previous cards. Instead, it states a "maximum clock speed" that the card could reach given enough thermal headroom. Once it reaches the thermal limit, which is exactly 94C on the R9 290X, it begins to throttle the clock speeds a bit to keep temperatures in check. When clock speeds go down, so does performance. Now, if clock speeds just go down a tiny bit, like 50MHz, performance won't suffer that much. However, Nvidia claims that when the R9 290X is set to its default "quiet" mode that clock speeds can go as low as 700MHz, and then stay in that neighborhood until the card cools down, resulting in reduced overall performance.
In our testing we did not experience a radical decline in clock speeds on the R9 290X. Sure, it fluctuates but generally stays above 900MHz. We even ran some tests to see how much our R9 290X press board would fluctuate, so we let the card get up to 94C and then ran Heaven 4.0 and recorded a score of 33.4 frames per second (we know the chart above shows 36fps). We then let the R9 290X run overnight, which was approximately 16 hours, in order to ensure the card was hot as Hades. We then ran the Heaven 4.0 test again, and the score was 33.6 frames per second, so it did not change over time despite being as hot as possible. We also examined the bar graph showing clock speed changes over that time period, and though there were small dips, it was still pretty consistent. These tests were performed with the card in its stock mode, which is "quiet" as the fan never goes above 40 percent. It's in this mode that you will see the most clock speed fluctuation, as in "Uber" mode with the fan running at about 50 percent, there is very little fluctuation since the card's temps are more under control.
This screenshot was taken after Heaven 4.0 had been running on the R9 290X for 16 hours. In this image you can see the GPU clock speed over time, the fan set to 40 percent (Quiet mode) and the temp of 94C. That is the R9 290X's standard operating temperature under load. Click the image to see it in full-resolution.
Here's the rub: Even though the card provided to us by AMD didn't exhibit drastic clock speed fluctuation, other news outlets are reporting that retail boards acquired through e-tailers are showing major fluctuations. This would indicate that the board provided to the press were "golden samples," or boards tested or configured to not exhibit the same behavior seen in retail boards. This is obviously a problem, for several reasons. The boards we receive should be *exactly* the same as retail boards, period. But in this instance something is amiss, either with the press boards or with the retail boards, at least according to sites like the
. AMD says the problem lies with the retail boards, and it's working on a driver fix that will "minimize this variance" according to the statement provided to the Tech Report. For what it's worth,
a site in Sweden
also obtained retail R9 290X boards and found the benchmark scores to be identical to those of the press board. We will be obtaining a retail R9 290X and will post our test results soon.
To Nvidia's credit, it specs its boards with a Base Clock that is guaranteed, and performance can only go up from there if you overclock. AMD, at least this time around, is doing the opposite by stating the maximum clock speed the card can achieve in ideal conditions, with performance only dropping from there. How much it drops is an area of debate currently, but just to be clear, in our testing we did not experience the drastic clock speed fluctuations reported in the retail cards, and by Nvidia. Even in our overnight test of the R9 290X we did not see a drop in performance.
With the release of the GTX 780 Ti Nvidia lays claim to the fastest single GPU in the world title once again. We haven't seen a card dominate the high-end proceedings like this in a while, probably since the GTX Titan was released actually. Not only is it fast, but like the other Kepler cards it's cool and quiet, two traits that have gained new appreciation this week as gamers consider the new Hawaii cards from AMD. Both of those cards represent very strong price-to-performance ratios, but neither of them run hot, and are noticeably louder than their Nvidia equivalents. We don't think the heat and noise are deal breakers, however.
Naturally, the GTX 780 Ti costs significantly more than the R9 290X, so we would expect it to outperform it by the same amount, and it certainly does. Barring some unforeseen new GPU from AMD it seems like Nvidia will remain the uncontested fastest GPU provider for the near future, at least until its new Maxwell cards come online sometime in 2014.
The only visible change to the now-familiar cooling apparatus is a smoke tint on the shroud and black lettering. We think it looks cool, so it's a shame you'll never see it once you install it.