I get a kick out of looking back at what I wrote about certain pieces of high-end hardware. When Nvidia launched the GeForce GTX 680, AMD was still asking something like $550 for the Radeon HD 7970, and the GK104-based board kicked it right in the tail. It was faster, cooler, quieter, and smaller than AMD’s flagship. I recommended the 680 without hesitation. And until Nvidia launched its almost-as-fast GeForce GTX 670 for even less, the 680 was a great choice.
Today, 7970s are down to $400 or so. Meanwhile, the GTX 680s are selling for roughly $460. What a reversal, right? After some serious driver work, AMD’s Radeon HD 7970 is notably faster than the GeForce board and it costs less. You have to put up with more noise and higher power use, but the Tahiti-based card also gives you great compute horsepower to match its gaming alacrity. So long as you stay away from multi-GPU configurations, the Radeon HD 7970 is a smart buy.
The next step up is going to cost you a cool grand. Be it the GeForce GTX Titan, the GeForce GTX 690, or AMD’s Radeon HD 7990, single-card performance doesn’t get any better than a Radeon HD 7970 unless you spend twice as much. And that’s where AMD and Nvidia lose a lot of gamers otherwise down to drop big bucks on graphics. Stepping up from $500 to $1,000 is rough.
GK110 Finds A New Home In GeForce GTX 780
GeForce GTX 780 is Nvidia’s attempt to do a little something about that gaping maw of a price delta between GTX 680 and the crazy-expensive stuff. Given its name, you might think the 780 centers on a new piece of silicon. But it’s really a derivative of GeForce GTX Titan and the gargantuan GK110 GPU.
Of course, the GK110 that Nvidia uses on GeForce GTX 780 is necessarily trimmed to keep it from showing up the potent Titan. We already know that a complete GK110 GPU plays host to 15 Streaming Multiprocessors, each with 192 CUDA cores and 16 texture units. GeForce GTX Titan pares the chip back to 14 SMXes, totaling 2,688 CUDA cores and 224 texture units. GeForce GTX 780 sees GK110 further cut down to 12 SMXes. The result is 2,304 CUDA cores and 192 texture units.
GK110 as it appears in GeForce GTX 780
Depending on the card you get, GeForce GTX 780’s 12 SMX blocks are either spread between four or five Graphics Processing Clusters. Composed of 7.1 billion transistors, GK110 is a massive chip. Manufacturing it isn’t easy. And along the way, different parts of it show up with defects. So, Nvidia can’t guarantee the exact configuration of each GeForce GTX 780’s GK110. It’ll only say that, across the GPU, 12 SMXes are enabled.
Nvidia’s incisions are effective enough in dictating performance that tweaks to the board’s clock rates are very minor. GeForce GTX 780 bears an 863 MHz base frequency, just like Titan. But its rated GPU Boost clock rate is 900 MHz, whereas Titan is officially spec’ed at 876 MHz.
GK110 retains its complete render back-end, including six ROP partitions able to output eight 32-bit pixels per clock, adding up to what the company calls 48 ROP units. Further, a sextet of 64-bit memory interfaces yield the same 384-bit aggregate pathway. But whereas Nvidia armed GeForce GTX Titan with 6 GB of GDDR5 memory, GTX 780 sports 3 GB operating at 1,502 MHz. Do the math and you get the same 288.4 GB/s of peak bandwidth.

Where GeForce GTX 780 veers away from Titan is in compute potential. You’ll remember from Nvidia GeForce GTX Titan 6 GB: GK110 On A Gaming Card that Nvidia’s single-GPU flagship includes a special driver setting that scales back clock rate in favor of running the chip’s double-precision floating-point units at full-speed. This makes the GeForce GTX Titan a viable option for developers seeking more compute performance than Nvidia’s other GPUs can muster (making it competitive with AMD’s Tahiti, in fact). This time around, you still get 64 FP64 CUDA cores per SMX. But because that driver setting isn’t exposed, double-precision performance drops back to 1/24 of the FP32 rate. Expect floating-point performance to trail Radeon HD 7970 then, as FP64 throughput lags.
| GeForce GTX Titan | GeForce GTX 690 | GeForce GTX 780 | GeForce GTX 680 | Radeon HD 7970 GHz Ed. | |
|---|---|---|---|---|---|
| Shaders | 2,688 | 2 x 1,536 | 2,304 | 1,536 | 2,048 |
| Texture Units | 224 | 2 x 128 | 192 | 128 | 128 |
| Full Color ROPs | 48 | 2 x 32 | 48 | 32 | 32 |
| Graphics Clock | 836 MHz | 915 MHz | 863 MHz | 1,006 MHz | 1,000 MHz |
| Texture Fillrate | 187.5 Gtex/s | 2 x 117.1 Gtex/s | 165.7 Gtex/s | 128.8 Gtex/s | 134.4 Gtex/s |
| Memory Clock | 1,502 MHZ | 1,502 MHz | 1,502 MHz | 1,502 MHz | 1,500 MHz |
| Memory Bus | 384-bit | 2 x 256-bit | 384-bit | 256-bit | 384-bit |
| Memory Bandwidth | 288.4 GB/s | 2 x 192.3 GB/s | 288.4 GB/s | 192.3 GB/s | 288 GB/s |
| Graphics RAM | 6 GB GDDR5 | 2 x 2 GB GDDR5 | 3 GB GDDR5 | 2 GB GDDR5 | 3 GB GDDR5 |
| Die Size | 551 mm2 | 2 x 294 mm2 | 551 mm2 | 294 mm2 | 365 mm2 |
| Transistors (Billion) | 7.1 | 2 x 3.54 | 7.1 | 3.54 | 4.31 |
| Process Technology | 28 nm | 28 nm | 28 nm | 28 nm | 28 nm |
| Power Connectors | 1 x 8-pin, 1 x 6-pin | 2 x 8-pin | 1 x 8-pin, 1 x 6-pin | 2 x 6-pin | 1 x 8-pin, 1 x 6-pin |
| Maximum Power | 250 W | 300 W | 250 W | 195 W | 250 W |
| Price (Street) | $1,000 | $1,000 | $650 | $460 | $450 |
Playing The Name Game
Perhaps you’re asking: Why call this card GeForce GTX 780 at all, then? It’s a derivative of Titan, based on the same Kepler architecture already prolific across the GeForce GTX 600 series. Nvidia did much the same thing with its 500 series, which built on the GeForce GTX 400’s Fermi architecture. “But the 500s were based on redesigned GPUs that improved performance, power, and, consequently, efficiency,” you rightly point out.

The most I could get from Nvidia was that it didn’t need to do this for the GeForce GTX 780, since the company only just started releasing desktop-oriented cards based on GK110. And there’s really not much room left in the 600 family to release newer, faster products. We’ll see how Nvidia fleshes out the performance and pricing of its GeForce GTX 700 line-up from here. But you can bet we’re going to expect notable performance improvements each step of the way to justify the naming. We also need continued (and compelling) competition from AMD to keep Nvidia’s pricing in check. Our best hope for that today is Radeon HD 7970.
I continue to be impressed by Nvidia’s industrial design. Sexy isn’t a word I typically ascribe to a piece of PC hardware, but it’s hard not to admire the GeForce GTX 690, Titan, and now the 780. Aesthetically, GeForce GTX 780 is almost identical to Titan, aside from the GTX 780 etched in the front of the card’s shroud. Consequently, everything I said about Titan in my February launch coverage applies here, too.

We’re looking at another 10.5”-long board, which, again, is half of an inch shorter than AMD’s Radeon HD 7970. And whereas the Tahiti-powered competition employs a cheaper-feeling plastic cover, GTX 780 sports a familiar aluminum shell surrounding a polycarbonate window that peers into the card’s heat sink. Unfortunately, the magnesium alloy fan housing we were treated to on GeForce GTX 690 is gone, just as it was with Titan.

Similar also is the centrifugal fan that effectively exhausts heated air from the 780’s rear I/O panel. This is a particularly big deal in multi-card configurations, since you don’t want waste heat getting recirculated back into your case, sabotaging your CPU overclock.

One thing Nvidia says it does improve on GeForce GTX 780 compared to Titan, however, is fan control. The company boasts that it developed a controller with an adaptive temperature filter and a thermally-targeted software algorithm that more intelligently maintains consistent fan speeds. On paper, the change is quite small—100-RPM fluctuations are tightened up to a roughly 20-RPM range. But that’s enough, purportedly, to bring noise down a few decibels compared to the GeForce GTX 680.
Oddly, no acoustic comparisons are drawn to Titan, which we assume means that noise is comparable between the two cards. This makes sense, given similar GPUs and industrial designs. Moreover, the GeForce GTX 780 and Titan share a 250 W TDP. So it’s hardly a surprise that they both have eight- and six-pin auxiliary power connectors, too.

Display connectivity also gets replicated from the GeForce GTX Titan to the 780. Two dual-link DVI outputs, HDMI, and DisplayPort enable as many as four simultaneous screens. You can use three in Surround and a fourth as a Windows desktop screen, just sort of hanging out. By far, triple-screen arrays are the most attractive with a card like this, though.
Nvidia says that the GeForce GTX 780 will completely replace the GeForce GTX 680 in its line-up and sell for $650. That’s an almost $200 premium for 780’s additional performance in games and a bit of acoustic dampening. Meanwhile, Radeon HD 7970 GHz Edition cards are selling for $450 and significantly outpacing the old GK104-based 680. Without a solution back down in the GeForce GTX 680’s price range, AMD is going to eat that space up (particularly with non-GHz Edition cards going for $400). It’s a pretty poorly-kept secret that more GeForce GTX 700-series cards are on their way, and those should help. The question for today is: does GeForce GTX 780’s speed warrant that greater-than 40% increase in price?
GeForce Experience
As PC enthusiasts, we appreciate the permutations of different quality- and performance-oriented settings available to us in the latest games. It’s actually validating to spend a bunch of money on a new graphics card and then start cranking up the graphics details to the highest levels. When there’s an option that pushes our hardware too far and remains unplayable, we’re irked knowing that the title can look that much better.
Dialing in the optimal settings isn’t easy, though. Some yield better visuals than others, while performance is impacted to varying degrees. GeForce Experience is Nvidia’s attempt to simplify choosing game settings by comparing your CPU, GPU, and resolution to a database of configurations. The other half of the utility makes sure your drivers stay up to date.
I still think that enthusiasts are going to resist running an additional piece of software to make these decisions for them. But more mainstream gamers inclined to install a title and fire it up immediately, without checking for drivers or tweaking its settings, naturally stand to gain quite a bit. Insofar as GeForce Experience helps those folks enjoy a more compelling experience, Nvidia’s software is good for PC gaming.
GeForce Experience detected all nine of the games installed on our test bed. Of course, none of them were still at their default settings, since they were set up for benchmarking. It’d be interesting, I thought, to see how Experience would change the options I dialed in myself.
For Tomb Raider, GeForce Experience wanted to disable TressFX, even though the GeForce GTX 780 averages in the mid-40 FPS range with it on. The utility wasn’t able to detect our Far Cry 3 configuration at all for some reason, though the settings it suggested were sufficiently high-end. In Skyrim, it wanted to turn off FXAA for some strange reason.
It’s very cool to get a set of screen shots for each game with descriptive boxes explaining what effect each setting has on image quality. From the nine examples I looked over, GeForce Experience did come close to the settings I would have selected. But it also biases to Nvidia-specific features like PhysX (which it cranks up to High in Borderlands 2) and against those introduced by AMD (including TressFX in Tomb Raider). Disabling FXAA in a game like Skyrim just doesn’t make sense given near-100 FPS average frame rates. It’ll likely be more important for enthusiasts to have GeForce Experience installed once Shield starts shipping; the Game Streaming feature appears to be enabled through Nvidia’s app.
Update, 5/28: Nvidia sent over some feedback indicating that TressFX is disabled by GeForce Experience using the configuration and resolution we selected in Tomb Raider because previous versions of the game achieved a significantly lower performance level. So, it seems there will be some delay between when drivers or patches dramatically improve frame rates, and when GeForce Experience takes those improvements into consideration. For now, I maintain that this is a mismatch in the settings an enthusiast would use compared to the automated system. In Skyrim, however, Nvidia clarifies that FXAA is not enabled by default because this feature is actually bad for image quality when used in conjunction with MSAA. Thus, the decision to keep FXAA turned off was deliberate, and not an oversight. Finally, the company is adamant it does not weigh its own technologies higher than the competition's. To illustrate, it points out that PhysX is disabled in Planetside 2, favoring other detail settings that contribute more to the experience. Specifically, "PhysX effects can improve the game play experience immensely (as seen in a game like Batman: Arkham City), but PhysX or TXAA are never given higher priority in GFE just because they’re Nvidia technologies." We appreciate Nvidia's feedback, which of course is the product of many more game titles than we have available here for testing. As GeForce Experience matures, we look forward to seeing how gamers at all levels react to it.
ShadowPlay: An Always-On DVR For Gaming
A long time ago, in a galaxy far away, I was pretty into WoW. Like, I raided four days a week, multiple hours a day. Any time our guild tackled a new boss, we’d record every attempt in the hopes we’d score a kill. When we did, one of us would post a video of it. Invariably, the people recording the video needed a decently high-end machine, Fraps, and plenty of storage space for all of those hours of wipes.
I don’t do that anymore. But when Nvidia talked to me about its new ShadowPlay feature, raiding is what instantly came to mind.

Enabled, ShadowPlay leverages the NVEnc fixed-function encoder built into Kepler-based GPUs and automatically records the last 20 minutes of game play. Or, you can manually control when ShadowPlay starts and stops doing its thing. This replaces software-based solutions like Fraps, which exact a more demanding load on your host processor.
As a quick refresher from our GeForce GTX 680 launch story, NVEnc is limited to H.264 encodes at up 4096x4096. ShadowPlay is not yet available, but Nvidia says that when shows up later this summer, it’ll do 1080p recordings at up to 30 FPS. I’d like to see it capable of higher resolutions, given earlier claims that the encoder can handle them in hardware.
GPU Boost 2.0
I didn’t have a chance to do a ton of testing with Nvidia’s second-generation GPU Boost technology in my GeForce GTX Titan story, but the same capabilities carry over to GeForce GTX 780. Here’s the breakdown:
GPU Boost is Nvidia’s mechanism for adapting the performance of its graphics cards based on the workloads they encounter. As you probably already know, games exact different demands on a GPU’s resources. Historically, clock rates had to be set with the worst-case scenario in mind. But, under “light” loads, performance ended up on the table. GPU Boost changes that by monitoring a number of different variables and adjusting clock rates up or down as the readings allow.
In its first iteration, GPU Boost operated within a defined power target—170 W in the case of Nvidia’s GeForce GTX 680. However, the company’s engineers figured out that they could safely exceed that power level, so long as the graphics processor’s temperature was low enough. Therefore, performance could be further optimized.
Practically, GPU Boost 2.0 is different only in that Nvidia is now speeding up its clock rate based on an 80-degree thermal target, rather than a power ceiling. That means you should see higher frequencies and voltages, up to 80 degrees, and within the fan profile you’re willing to tolerate (setting a higher fan speed pushes temperatures lower, yielding more benefit from GPU Boost). It still reacts within roughly 100 ms, so there’s plenty of room for Nvidia to make this feature more responsive in future implementations.
Of course, thermally-dependent adjustments do complicate performance testing more than the first version of GPU Boost. Anything able to nudge GK110’s temperature up or down alters the chip’s clock rate. It’s consequently difficult to achieve consistency from one benchmark run to the next. In a lab setting, the best you can hope for is a steady ambient temperature.
In addition to what I wrote for Titan, it should be noted that you can adjust the thermal target higher. So, for example, if you want GeForce GTX 780 to modulate clock rate and voltage based on an 85- or 90-degree ceiling, that’s a configurable setting.
Eager to keep GK110 as far away from your upper bound as possible? The 780’s fan curve is completely adjustable, allowing you to specify duty cycle over temperature.
Troubleshooting Overclocking
Back when Nvidia briefed me on GeForce GTX Titan, company reps showed me an internal tool able to read the status of various sensors, which made it possible to diagnose problematic behavior. If an overclock was pushing GK110’s temperature too high, causing a throttle response, it’d log that information.

The company now enables that functionality in apps like Precision X, triggering a “reasons” flag when certain boundaries are crossed, preventing an effective overclock. This is very cool; you’re no longer left guessing about bottlenecks. Also, there’s an OV max limit readout that lets you know if you’re pushing the GPU’s absolute peak voltage. If this flag pops, Nvidia says you risk frying your card. Consider that a good place to back off your overclocking effort.
| Test Hardware | |
|---|---|
| Processors | Intel Core i7-3770K (Ivy Bridge) 3.5 GHz at 4.0 GHz (40 * 100 MHz), LGA 1155, 8 MB Shared L3, Hyper-Threading enabled, Power-savings enabled |
| Motherboard | Gigabyte Z77X-UD5H (LGA 1155) Z77 Express Chipset, BIOS F15q |
| Memory | G.Skill 16 GB (4 x 4 GB) DDR3-1600, F3-12800CL9Q2-32GBZL @ 9-9-9-24 and 1.5 V |
| Hard Drive | Crucial m4 SSD 256 GB SATA 6Gb/s |
| Graphics | Nvidia GeForce GTX 780 3 GB |
| AMD Radeon HD 7990 6 GB | |
| AMD Radeon HD 7970 GHz Edition 3 GB | |
| Nvidia GeForce GTX 580 1.5 GB | |
| Nvidia GeForce GTX 690 4 GB | |
| Nvidia GeForce GTX 680 2 GB | |
| Nvidia GeForce GTX Titan 6 GB | |
| Power Supply | Cooler Master UCP-1000 W |
| System Software And Drivers | |
| Operating System | Windows 8 Professional 64-bit |
| DirectX | DirectX 11 |
| Graphics Driver | AMD Catalyst 13.5 (Beta 2) |
| Nvidia GeForce Release 320.00 | |
| Nvidia GeForce Release 320.18 (for GeForce GTX 780) | |
Getting Frame Time Variance Right
Astute readers will notice that the numbers on the following page (and those thereafter) are quite a bit more conservative than the same page in my Radeon HD 7990 review, and there is a reason for this. We were previously reporting the raw and real-world frame rates, and then showing you the frame time variance data with runt and dropped frames still included. The thing is, if that’s not what you experience, it isn’t fair to then point to the raw frame time latencies and hammer AMD on them.
This is why we’re now giving you the more practical frame rates over time, along with frame rate variance numbers that match. The outcome is far less exaggerated, though still very telling in terms of the games where AMD struggles.
| Benchmarks And Settings | |
|---|---|
| Battlefield 3 | Ultra Quality Preset, v-sync off, 2560x1440, DirectX 11, Going Hunting, 90-Second playback, FCAT |
| Far Cry 3 | Ultra Quality Preset, DirectX 11, v-sync off, 2560x1440, Custom Run-Through, 50-Second playback, FCAT |
| Borderlands 2 | Highest-Quality Settings, PhysX Low, 16x Anisotropic Filtering, 2560x1440, Custom Run-Through, FCAT |
| Hitman: Absolution | Ultra Quality Preset, MSAA Off, 2560x1440, Built-In Benchmark Sequence, FCAT |
| The Elder Scrolls V: Skyrim | Ultra Quality Preset, FXAA Enabled, 2560x1440, Custom Run-Through, 25-Second playback, FCAT |
| BioShock Infinite | Ultra Quality Settings, DirectX 11, Diffusion Depth of Field, 2560x1440, Built-in Benchmark Sequence, FCAT |
| Crysis 3 | Very High System Spec, MSAA: Low (2x), High Texture Resolution, 2560x1440, Custom Run-Through, 60-Second Sequence, FCAT |
| Tomb Raider | Ultimate Quality Preset, FXAA Enabled, 16x Anisotropic Filtering, TressFX Hair, 2560x1440, Custom Run-Through, 45-Second Sequence, FCAT |
| LuxMark 2.0 | 64-bit Binary, Version 2.0, Sala Scene |
| SiSoftware Sandra 2013 Professional | Sandra Tech Support (Engineer) 2013.SP1, Cryptography, Financial Analysis Performance |
I explained a lot of the methodology we’re using in AMD Radeon HD 7990: Eight Games And A Beastly Card For $1,000. But again, our GeForce GTX 780 review is going to have a lot of data to comb over.

As we might expect, the GeForce GTX 780 falls in just behind Nvidia’s GeForce GTX Titan. In that context, the $650 card offers quite a bit more value at its price point than the flagship. However, AMD’s $450 Radeon HD 7970 GHz Edition board isn’t far behind either, outpacing the more expensive GeForce GTX 680.
What’s going on with the Radeon HD 7990? Well, if we were to believe our hardware-based numbers, AMD’s dual-GPU beast would be the winner in this benchmark. But after we remove dropped and runt frames from the analysis, one of the card’s GPUs appears to be doing very little useful work, yielding a frame rate you can actually see just above a single Radeon HD 7970.

Here’s what that average frame rate chart looks like plotted over time. Although we illustrated hardware and practical performance last time, we’re cleaning the data up as much as possible, giving you just the numbers that matter: practical frame rate.
Nvidia’s GeForce GTX 780, in red, trails the Titan ever so slightly, just ahead of the Radeon HD 7970, but far more consistent than the Radeon HD 7990.

Our attempts to simplify result in a less complex frame time variance chart, too. As a reminder, this is a reflection of the latency between consecutive frames, not the absolute time it takes to render a frame. Rather than average, 75th, and 95th percentile numbers, we have the average and 99th (worst-case) results.
On average, none of those cards are terrible in Battlefield 3. Based on the real-world experiential stuff I’ve done with gamers, it seems like 5 ms of difference is the point where folks start consistently picking the card with lower variance over the other. Conversely, we see that the Radeon HD 7990 can get pretty bad when we look at the entire run, minus the worst 1%. We already know this, though—until AMD can work out its multi-GPU pacing issues, examples like this should give gamers a moment of pause about spending $1,000 on a Radeon HD 7990.

GeForce GTX 780 slots in behind Titan in BioShock Infinite, and ahead of the $450 Radeon HD 7970 GHz Edition. The card’s advantage over its predecessor is significant—almost 60%.

Charting the practical frame rate over time demonstrates performance that generally stays above 50 FPS, and right under the GeForce GTX Titan.

When it comes to average frame time variance, AMD’s Radeon HD 7990 fares worst, though calculating worst-case latency isn’t kind to the GeForce GTX 580 either. The two-generation-old board spends much of its time under 30 FPS, so we aren’t considering it playable at these settings anyway.

A 15% advantage over GeForce GTX 680 is perhaps less impressive than the fact that GeForce GTX 780 very nearly matches the pricier Titan board. Interesting also is that the $1,000 Titan is just about as fast as the Radeon HD 7990. Both of those pricey products fall just slightly under the GeForce GTX 690.

The tight clump of lines in our practical frame rate over time chart is indicative of a title that isn’t particularly difficult for these high-end cards to handle.

A platform-oriented bottleneck relaxes the load on graphics processing, yielding relatively low frame-to-frame latency. All of the single-chip Kepler-based board perform fairly similarly.

If you want the skinny on why a Radeon HD 7990 is simultaneously well-behaved and a poor performer, check out my analysis in the review of that card. Today we’re focusing on the GeForce GTX 780, which is again almost as fast as a GeForce GTX Titan. A 31% speed-up compared to the GeForce GTX 680 isn’t bad either, even if that’s not as high as the 41% Nvidia plans to charge for GeForce GTX 780.

At these settings, a GeForce GTX 690 is really your best bet. The GTX 780 and Titan do trade blows, though dipping under 30 FPS would compel us to scale back the detail settings.
The impact of mediocre performance in Crysis 3 is particularly pointed when it comes to running and gunning. Low frame rates cause significant lag, making it hard to shoot accurately. I really needed multi-GPU configurations from Nvidia in order to make the game run well at 2560x1440.

The average variance of one Radeon HD 7970 or the dual-GPU Radeon HD 7990 is almost identical, though the 7990’s 99th percentile number is significantly worse.
To that end, while a GeForce GTX 690 tends to do really well on average, its 99th percentile latency is also bad.
Nvidia’s GeForce GTX 780 does well from one frame to the next, despite the lower-than-ideal performance level at this combination of settings.

GeForce GTX 780 manages to outperform Titan by a hair in Far Cry 3. Implausible though this may seem, remember that GPU Boost opportunistically improves performance. We keep the lab at a constant temperature and run these cards through the same benchmarks in the same order, but it’s impossible to recreate the same conditions between tests. Simply, there are going to be cases where GeForce GTX 780 and Titan behave similarly.

Clearly a graphics-bound title, Far Cry 3 puts plenty of space between most cards. The only two that overlap are the GK110-based boards.
Based on the average frame rate results, AMD’s Radeon HD 7990 is faster than the GeForce GTX 780 and Titan. However, it’s problematic that the dual-Tahiti-based card dips into the 20-FPS range. Maintaining from 35 to 40 FPS is preferable, even if Nvidia’s top-end single-GPU cards achieve slightly lower averages.

The frame rate over time graph showed AMD’s Radeon HD 7990 bouncing up and down. Large swings in instantaneous frame rate made it pretty easy to predict the outcome of this chart. On average, the 7990’s variance from one frame to the next is worse than a GeForce GTX 580’s worst-case variance. In comparison, the GeForce GTX 780 is very consistent, falling between the GeForce GTX 690 and Titan.

GeForce GTX 780 again manages performance on par with the Titan board. Both cards are overwhelmed by the much less expensive Radeon HD 7970 GHz Edition, though. We can see something is holding back Nvidia’s hardware—even the GeForce GTX 690 can’t break past 55 FPS.

The three fastest boards from Nvidia all track similarly when we chart frame rate over time. A performance floor around 45 FPS isn’t really an issue in this case, though it might bother some of the folks with $1,000 graphics cards that a $450 Radeon HD 7970 GHz Edition is quicker for no discernable reason.

The only real notable result comes from AMD’s Radeon HD 7990, which dominates the average frame rate chart, but also exhibits the greatest degree of variance from one frame to the next. Close to 9 ms isn’t severe, but again, we’ve seen as few as 5 ms sway a group of gamers in favor of the smoother delivery.

We commonly point and laugh at Skyrim due to its light graphics load, which you wouldn’t think would make it a very meaningful GPU benchmark. But look—there’s plenty of performance difference between a GeForce GTX 780 and 680 (36%).

All of these cards maintain at least 50 FPS through our test run at 2560x1440. And yet, the game still scales really well.

On average, all seven graphics configurations we’re testing maintain sub-1 ms frame time variance. This means the action you see on-screen is smooth and consistent. Worst-case, though, the Radeon HD 7990 approaches 10 ms between frames, while the Radeon HD 7970 achieves an amazing .6 ms. All of Nvidia’s cards show up in between.

The $650 GeForce GTX 780 is a little faster than the $450 Radeon HD 7970 GHz Edition in Tomb Raider. In turn, the Radeon outmaneuvers the GeForce GTX 680, which costs about the same. This is a great showing for the Tahiti-based board.

The Radeon HD 7990 exhibits the highest highs and comes close to the lowest lows—anyone want to guess what the variance chart is going to look like? We see Nvidia’s new GeForce GTX 780 tracking uncannily close to the Radeon HD 7970 GHz Edition.

We get solid average frame time variance numbers from all of the tested cards. But as we predicted, the Radeon HD 7990 runs into trouble when things get bad, nearing 20 ms, even after we disregard the worst 1% of latencies.

This is where the expensive GeForce GTX Titan and 780 earn their stripes. When it comes to pure performance, AMD’s Radeon HD 7970 GHz Edition is a strong contender for $200 less than what Nvidia is asking. But that only applies when you use one 7970 on its own. Start pairing cards up, and the average frame rates you see on-screen (after factoring out runt and dropped frames) diverge dramatically.
Two GeForce GTX 780s run $1,300. Two Titans go for $2,000. The 690 is a $1,000 card. And a pair of 680s run $920. If you calculate what you pay for FPS in Battlefield 3 at 2560x1440, the 780s and 680s aren’t all that far apart (under $11 and $10/FPS, respectively), while the dual Titans cost quite a bit more (over $15/FPS).

Two GK110-based boards put you in an entirely different class of performance. And consider that, for $50 less than two GeForce GTX Titans, you could actually have three 780s.
That frame rate over time line doesn’t bode well for the frame time variance of AMD’s Radeon HD 7990.

This isn’t as bad as I was expecting from AMD’s cards, though remember the runts and drops are already factored out. That’s why their average frame rate and frame rate over time results are so disappointing.

This is another strong showing from the GeForce GTX 780s, which serve up much greater value than the Titan cards.


We already know from our Radeon HD 7990 review that gamers playing on a system with AMD’s dual-Tahiti flagship and Nvidia’s GeForce GTX 690 universally picked the twin-GK104-based card as delivering the smoother experience in BioShock Infinite. That 99th percentile difference of 6 ms between the two boards is far more significant than we could have guessed prior to our blind-testing experiment.

Based on the average numbers above and frame rate over time chart below, the strange results in Borderlands 2 could be a platform bottleneck issue, though we’d expect the GeForce GTX 680s in SLI to be similarly affected. Nevertheless, a look back at our raw FCAT data indeed shows the 780s averaging lower performance.


Despite the lower frame rates, two GeForce GTX 780s deliver consistent pacing, bested only by a pair of Titans, by our measurements.

We’re back to blistering frame rates—two GeForce GTX 780s come within 4 FPS of the Titans, on average.


Those 99th percentile results are precisely why it can be hard to run and gun through Crysis using the most demanding settings. Play this title on a set of cards that runs smoothly and you’ll never be able to go back.

One GeForce GTX 780 gets you 38 FPS in our Far Cry 3 benchmark, and two get you 68 FPS. That’s the difference between marginal performance and completely smooth at 2560x1440.

Frame rates over time show us that the 780s stay above 64 FPS through our entire test run. The 680s and 690 keep above 50 FPS, while the AMD cards struggle, even with two GPUs working cooperatively.

The GeForce GTX 780s maintain excellent pacing between consecutive frames, giving us a worst-case of 3.6 ms. Compared to the Tahiti-based boards that approach 30 ms between frames, the difference is significant.

The same bottleneck that stopped one GeForce GTX 780 kicks two of them into last place, within .5 FPS of three other Nvidia-based configurations.


The AMD combos achieve higher frame rates in Hitman’s canned benchmark, but they continue to demonstrate notably higher latencies between consecutive frames.

Given the platform-based bottleneck imposed by Skyrim, there’s little reason to add a second high-end card for this game alone. In fact, only the GeForce GTX 680 and Radeon HD 7970 pick up any extra speed—and what they gain isn’t worth what you pay.


Minimal scaling aside, the AMD cards don’t see their frame-to-frame latencies narrow. In fact, whereas a single Radeon HD 7970 demonstrates a worst-case of .6 ms, two working together spike to 8.6 ms.

One GeForce GTX 780 manages 44.66 FPS in our Tomb Raider benchmark; two achieve 85.52 FPS on average.


We get one last reminder that high frame rates don’t necessarily guarantee consistent pacing. Although the Radeon HD 7990 delivers an admirable average frame time variance, the 99th percentile result shows that you’ll experience pauses of up to 20 ms between frames.
Heat
The reference cooler's fan curve is quite conservative. For example, the cooler only begins to spin faster at around 60°C, finally kicking into high gear upwards of 75°C. A direct comparison between the GeForce GTX Titan and 780 reveals the latter’s maximum fan speed to be slightly lower. In our gaming loop, Nvidia’s new card runs about two degrees cooler than the Titan, resulting in a slightly lower fan speed and a more gradually rising fan speed curve. This is due to the aforementioned refinements to the card's fan controller.


Despite the minor differences already noted, the two cards’ fan speed and temperature graphs converge more and more as time goes by, although they never actually intersect. That’s surprising, considering the surgical changes Nvidia made to its GeForce GTX 780. Let’s keep that first graph showing the temperatures in mind when as we move on to the next page. Specifically, we want to check the correlation between the thermal target of 80°C and power consumption.
Noise
But first, let’s take a closer look at the new card’s noise output, with the GeForce GTX Titan serving as our reference point again:

The GeForce GTX 780 enjoys a small advantage in each of our scenarios, although the differences are too small to be noticeable. In light of this, you could even consider the results close enough to be a tie. We’ll try to explain why this is so on the next page.
Clock Rates and Thermal Limits
We already covered GPU Boost 2.0 and, small though they may be, the subtle changes Nvidia made to GeForce GTX 780 compared to Titan. Specifically, we see that both cards start limiting core clock rates once the GPU reaches 60°C, dialing back the performance gains attributable to GPU Boost a bit. This becomes more pronounced on the GeForce GTX 780 as its core temperature rises, though the 780 exhibits much greater and more varied clock speed fluctuations, while the GTX Titan is basically limited to three clock levels. While GeForce GTX Titan practically loses the ability to boost its clock speeds as soon as it reaches its thermal limit, the graph still shows 780’s core frequency spiking upward consistently.

Power Consumption
In less demanding applications (including games), the GeForce GTX 780 uses slightly less power than Titan. Although both cards bear the same TDP, this is still somewhat surprising. You'd think that the pared-back hardware would be less power-hungry. The difference are in-line with what we're used to seeing from two similar cards with different amounts of RAM. In other words, it seems that the deactivated hardware blocks on the GTX 780 are still drawing power.




Also curious, the GeForce GTX 780 appears to use more power under full load than the Titan until it reaches its thermal limit. Meanwhile, the bigger card runs into its thermal limit sooner, while still offering more performance. One explanation is that the Titan operates closer to the GK110 GPU’s sweet spot, while the GTX 780 relies on higher clock speeds for its performance.

As long as the cards stay within their predefined thermal limits, they can hit power peaks beyond what their nominal TDP would allow. In practice, you will see these only rarely and very briefly at that. Still, don’t forget to take them into account and pick your power supply accordingly.

Effects of the Thermal Limit
Lastly, let’s look at what happens when both cards hit their thermal limits after being under load for a while. The GeForce GTX 780’s power consumption drops from 245 to 232 W, while the GeForce GTX Titan only dips by 2 W, from 238 to 236 W. This is another example of how much more headroom GPU Boost 2.0 provides, extracting extra performance as long as the GPU remains cool enough.

Synthetic Benchmarks
Unigine’s Heaven and Sanctuary benchmarks show us how cards perform when running demanding features from current gaming titles in OpenGL. Additionally, since none of the graphics drivers contain any optimizations for the OpenGL versions of these benchmarks, you could even say it’s a fairer comparison than using the super-optimized DirectX versions.


Maya 2013
If you’re not a fan of the DirectX 11-accelerated Viewport 2.0, Maya also continues to offer OpenGL support. The benchmark run we picked to represent performance also shows us that, generally, the Nvidia cards perform very similarly. That’s hardly a surprise, since the drivers for the consumer cards don’t contain any of the required optimizations.

SolidWorks
Since newer versions of this application only complete benchmark runs on professional graphics cards with validated drivers, we fall back on the older version that is part of the SPECviewperf11 benchmark suite.

EnSight
The same applies to EnSight, which is why we employ an older version here as well.

Aside from the two synthetic benchmarks, Nvidia's GeForce GTX 780 doesn't really outpace or fall behind the 680 by a meaningful margin. As expected, it consistently places just behind the GeForce GTX Titan.
AutoCAD 2013
Since AutoCAD 2013 is based on DirectX, we rely on the Cadalyst suite’s 2D and 3D performance index to gauge performance. In 2D mode, differences between the cards are marginal, so it doesn’t really matter which model you choose.

The 3D performance index paints a slightly different picture, with the GeForce GTX 780 just behind the Titan, while also enjoying a comfortable lead over the GeForce GTX 680.

Autodesk Inventor 2013
Inventor is another DirectX-based application. We're using the popular benchmark that rotates 1,000 cubes on-screen to test the cards’ limits. The tables turn for AMD. While the Radeon HD 7970 GHz Edition fell behind slightly in AutoCAD, it leads the pack of GeForce cards this time around, though the differences aren't huge.

CUDA Rendering
Since CUDA is a proprietary Nvidia technology, the Radeon HD 7970 GHz has to sit out the following four benchmarks. Despite its provenance as a GK110-based card, the GeForce GTX 780 always comes in behind the older 580 when it comes to pure rendering. The Octane benchmark is the sole exception. Meanwhile, the pricier Titan clearly dominates the field, while the GeForce GTX 680 finds itself relegated to last place just as consistently.



Synthetic: FluidMark
While the GeForce GTX 680 is barely able to keep the Titan at bay, the GTX 780 pulls ahead of both cards. We’re chalking this one up to a driver issue, though it's still weird.

We chose not to rely on the typical computing tests this time around, since we're specifically trying to isolate and compare single- and double-precision math performance. After all, like GeForce GTX Titan, the 780 is a descendant of the powerful Tesla cards. So, we're using metrics that let us test both levels of precision.
Financial Analysis Performance (Float/FP32)
Remarkably, the GeForce GTX 780 ends up closer to the 680 than the Titan.


Folding@Home (FP32)
The Folding@Home benchmark puts Nvidia's GeForce GTX 780 exactly halfway between the GeForce GTX Titan and GeForce GTX 680.


Looking at these results, two things become clear. First, Nvidia regrettably still appears to treat OpenCL as a second-class API. Second, the F@H benchmark proves that porting a CUDA application to OpenCL is not as trivial as it may sound. While the GeForce cards dominate when it comes to protein folding with the explicit solvent, AMD’s Radeon HD 7970 GHz Edition comes out on top in financial analysis performance.
Of course, it’s easy to find other benchmarks that show either one or the other company in the lead, but that’s not what we're after. Rather, we want to show how our field performs in these same tests switching to double-precision.
Although the GeForce GTX 780 shares large parts of its DNA with GeForce GTX Titan, the newer card’s driver does not offer the same option to speed up double-precision performance (at the cost of frequency). Consider this a matter of market segmentation by Nvidia, preventing the 780 from becoming a super-cheap development board for compute apps.
Financial Analysis Performance (FP64)
The Monte Carlo pricing test leaves no doubt that Nvidia purposely (and artificially) dumbs down the GeForce GTX 780’s FP64 capabilities in order to give Titan some breathing space.


Folding@Home (FP64)
With double-precision activated, the 780 retains its lead over the 680, but drops behind the aging GeForce GTX 580 in the Folding@Home benchmark.


Anyone who hoped that the GeForce GTX 780 would be a cheaper option for scientific computing is going to be disappointed. While this is an understandable move, it also wasn't necessary.
GeForce GTX Titan is a lot like Intel’s Core i7-3970X—a ridiculously fast piece of hardware sitting atop of a stack of alternatives that make a lot more sense. The GeForce GTX 780 is akin to Core i7-3930K. It’s the option we’d recommend to more savvy power users. Almost every bit as fast, it costs a lot less and sacrifices very little of the flagship’s feature set (FP64 performance the biggest loss).
But we’d be remiss if we didn’t point out the more value-oriented offering able to satisfy a majority of enthusiasts: Radeon HD 7970 GHz Edition at $450. If you average the performance of our eight benchmarks and then calculate what you pay for every frame per second, AMD’s single-GPU flagship runs $8.38/FPS. The GeForce GTX 780 lands at $10.73/FPS. The Tahiti-based board also maintains a massive advantage in compute-oriented workloads. And it still includes Tomb Raider, BioShock, Far Cry 3: Blood Dragon, and Crysis 3. That’s a killer bundle. When performance per dollar is your only consideration in a high-end graphics card, AMD comes away looking pretty good.
Conversely, the GeForce GTX 780 is faster in absolute terms, even though you pay more for every drop of extra speed. It’s quieter than 7970, it uses less power, it includes a number of tuner-friendly tools, and Nvidia has a really cool feature in ShadowPlay (too bad it isn’t available to test yet). The 780 is a much better-looking board, too. But an asking price of $650 is only a relief to someone who was about a pay a grand for Titan. By all other accounts, that’s still a big flippin’ number.
This isn’t the generational jump you see when a company updates its architecture on a smaller process node and hits you with more speed at a familiar price point. GeForce GTX 780 is a derivative of existing technology that drops its shoulder and charges its way into a new segment. Is it worth more than GeForce GTX 680 or Radeon HD 7970 GHz Edition? Absolutely. But I would have rather seen the 780 at $550 or $600.
When it comes to multi-card configurations, there’s a lot more to laud. For the price of two GeForce GTX Titans, you could have three 780s. For $300 more than one Titan, you can get two 780s. If you’re gaming at 2560x1440 or 5760x1080, a couple of GK104- or GK110-based boards will help maintain playable frame rates using the most demanding detail settings. A couple of Radeon HD 7970s or a Radeon HD 7990 might turn out decent benchmark results, but their frame pacing issues are noticeable enough to earn a panel of gamers' disdain in A/B testing next to a GeForce GTX 690. Two GeForce GTX 780s are a winning combination, if you have the means.
The GeForce GTX 780 is a sexy piece of graphics hardware built on top of an impossibly-complex 7.1 billion-transistor GPU. It’s very fast, very quiet, and includes several other attractive features. But, I’m going to wait a week before deciding what I’d spend my money on in the high-end graphics market. You’d be wise to do the same…





