Nvidia GeForce GTX 780

When Nvidia launched the GeForce GTX Titan in February, it set a new record for fastest single-GPU graphics card performance—and tagged the card with a correspondingly high price tag. At $1,000, the 7.1 billion transistor GPU was out of reach of all but the richest of gamers. Today, Nvidia is debuting the Nvidia GeForce GTX 780, a Titan-derived card. It offers most of the GTX Titan's performance but at almost half its price tag. All this earns the Nvidia GeForce GTX 780 our Editors' Choice.

Nvidia has trimmed a number of features to bring the price tag down from $1,000 to $649. The GTX 780 has 2,304 cores compared with Titan's 2,688 cores, a drop of approximately 15%. Texture mapping units are also somewhat lower, at 192, down from 224. The total number of raster operators (ROPs), however, is still the same, at 48. The chip's core clock is in the same ballpark as the Titan's (863MHz base, 902MHz boost, compared to the Titan's 836MHz base, 875MHz boost). What this means is that the GTX 780 has most of the Titan's raw horsepower, as well as the Titan's 384-bit memory bus and 6Gbps of DDR5 bandwidth.

The total amount of RAM, however, has been trimmed, down to "just" 3GB. This simply isn't much of a problem given that most games are still programmed in 32-bit and are therefore incapable of using more than 4GB of GPU memory in any case. The other casualty of the step-down includes full speed double-precision floating point (like previous GeForce products, the GTX 780 will execute DP code at 1/24th of native clock speed. Hyper-Q and Dynamic Parallelism, two of the high-end features from GK110, remain enabled in the GTX 780, and the chip still supports Compute Capability 3.5 rather than GK104's 3.0. This can matter to certain real-world performance tasks, including Bitcoin mining, though GK110-based cards are still far behind AMD in that regard.

Compared to the Nvidia GeForce GTX 680, the GTX 780 is a marked step forward. The new card has 50% more cores, 50% more texture mapping units, and 50% more render outputs than the older GK104-based solution. This is significantly offset by the lower clock speed; the GTX 680 ran at 1,006MHz standard with a 1,058MHz boost clock. That's a 15% difference between the GTX 680 and GTX 780, which ought to leave the newer card 25% to 35% faster than last year's model. The base memory configuration has also been bumped up to 3GB from the GTX 680's 2GB. That matches AMD's HD Radeon 7970, and it's a solid amount. Presumably we'll see a handful of GTX 780 cards pushing back to 6GB, the same way some vendors have built 4GB GTX 680 GPUs.

Our performance comparisons were done using an Intel 3770K Ivy Bridge CPU, 16GB of DDR3-1600, a 256GB OCZ Vector solid-state drive, and a 27-inch Asus VG278HE monitor at 1,920-by-1,080. The GTX 690 performance was simulated using a pair of GTX 680s in SLI; tests have demonstrated that the performance delta between the two configurations is essentially nil. All of our tests were run at 1,920-by-1,080 resolution with maximum details set. Multisampled antialiasing was activated when available and turned up to 8x if possible.

The GTX 780's speed boost over the GTX 680 varies depending on the game. In Civilization V's Late Game View, the GTX 780 was 12% faster than GTX 680 (100 frames per seconds vs. 89fps). In Metro 2033, that advantage jumped to 43% (45fps vs. 31.3fps). In Metro: Last Light, Metro 2033's just-released sequel, the GTX 780 was 22% faster, at 38fps vs. 31fps. In and BioShock Infinite, the GTX 780 is 16% and 11% faster than the GTX 680, (82fps vs. 70fps and 48fps vs. 43fps). It was also 30% faster in (110fps vs. 84fps).

If you intend to game at 1,920-by-1,080 resolution, the difference between the GTX 780 and the GTX Titan is often tiny. The GTX Titan was 6% faster in Metro 2033 (48fps vs 45fps), 7% faster in BioShock Infinite (88fps vs 82fps) and 3% faster in Civilization V (103fps vs. 100fps).

Performance against a single AMD Radeon 7970 was a bit more interesting. In games like Shogun 2, the GTX 780 torched the AMD card, 69fps vs. 38.5fps. In Metro 2033, however, the AMD 7970 was actually faster, at 56fps vs. 45fps for the GTX 780. Hitman: Absolution was another strong win for AMD; the Radeon 7970's 72fps beat out all the GeForce cards. Games like Arkham City and BioShock Infinite favor Nvidia, with the GTX 780 coming in at 82fps vs. the Radeon 7970's 68fps.

The final performance aspect we want to discuss is the topic of frame latencies, or the question of long it takes to draw each successive frame. Historically, this has been more of an issue for AMD than Nvidia, and it's a bigger problem with dual-GPUs than when single cards are tested. This, however, isn't an absolute—higher frame rates can still matter, and AMD's performance in this area is improving. Here's an example of GPU frame latencies in Metro: Last Light, as graphed by the open-source FRAFS tool, using FRAPS data. Shown here are the GTX 680, GTX 780, and the dual-GPU Radeon 7990 (Click on the images to enlarge).

Nvidia GeForce GTX 680:

Nvidia GeForce GTX 780:

AMD Radeon HD 7990

In this case, the dual GPU was noticeably smoother, thanks to fewer frame latencies that break the 33.3ms (30fps) barrier. As we've discussed at ExtremeTech [[link here]], there are other games where this isn't true—Shogun 2: Total War, is a title where AMD still noticeably lags Nvidia.

If you were looking at buying a Titan card for $1,000, the Nvidia GeForce GTX 780 for $649 is probably a great deal. The performance gap between the GTX 780 and GTX Titan is quite small at 1,920-by-1,080, and while we expect it would grow modestly at 2560x1600, it's not going to suddenly skyrocket. This makes the GTX 780 a great option for someone was already eyeing a GTX Titan and now wants to save a few hundred bucks. In fact, the only users who will seriously benefit from a Titan as compared to a GTX 780 are the handful of gamers looking to build multi-monitor rigs (at which point you probably want two Titan's, not just one), or anyone who had a serious use for the GTX Titan's full-speed double-precision floating point.

The GTX 780 is clearly faster than the GTX 680, but it commands a hefty price premium. In our suite of tests, the GTX 780 was, on average, 1.21x faster than the GTX 680, while costing 1.45x more. That's not unusual at these price brackets and there's definitely a market for these kind of products, but if you're searching for the best price/performance point on the market, the GTX 780 isn't it.

Comparisons against AMD products are even more lopsided. The non-GHz edition of the Radeon 7970 is available for $379 online, while the GHz flavor starts at $449. With the HD 7950 as low as $289, a pair of them would run $578—with four free games tossed in to boot, courtesy of AMD's "Never Settle" program. This is where the frame latency issue raises its head, however—if AMD's dual-GPU solutions could be absolutely counted on for delivering smooth playback, we'd say a pair of HD 7950s would likely trump the single GTX 780. With AMD making progress but not quite there, yet, this is less of a certainty.

High-end cards have always been about luxury as much as performance—and if you've been looking for a mini-Titan at a better price, you're going to be happy with the GTX 780. We give this high-end graphics card our Editors' Choice nod.