NVIDIA's GeForce GTX 780 Ti successfully captures the single GPU performance crown back from AMD. The card delivers impressive performance thanks to the GK110 GPU which has all its 2,880 shaders enabled. Averaged over all our benchmarks, performance is almost 10% higher than the GTX Titan and 8% higher than the R9 290X in its noisy "Uber" mode. The performance gap is 13% versus the R9 290X in quiet mode, or the R9 290, and the card even manages these performance levels with minimal throttling. Some investigation yielded an average clock rate of 993 MHz, which means that, thanks to NVIDIA GPU Boost 2.0, the card always stays well above NVIDIA's guaranteed 876 MHz the card will never go below. On AMD's R9 290 Series, we've seen major clock drops from time to time, depending on the heat situation, dropping the card below 700 MHz in some cases, though it is officially specified to run at "up to 1000 MHz."
During our testing, we also found the GeForce GTX 780 Ti to have fantastic overclocking potential. We managed to increase GPU base frequency by almost 30% (boost adds more clock on top of that), quite an impressive feat very uncommon to high-end reference designs. Unlike the R9 290X in quiet mode, overclocking also translated into real performance gains of 18.4%, which means the card does throttle, but not by much. Memory overclocks well, too, thanks to the fast Hynix 7 Gbps GDDR5 chips used. Even the GTX 780 Ti with its 384-bit bus at stock has more memory bandwidth than the R9 290X with 512-bit (336 GB/s vs. 320 GB/s). While 3 GB doesn't seem like a lot of memory, I am convinced it is enough for all games at all resolutions, and in the forseeable future. My guess is that we'll also see board partners release additional SKUs with 6 GB to cater to buyers who think they absolutely need more memory (even though they don't).
NVIDIA is using their Titan-style cooler on the GTX 780 Ti. I really like this cooler; it not only cools well, but also looks and (literally) feels nice thanks to the metal surfaces. It seems as though NVIDIA carefully tweaked their power/clock/cooling design to keep noise, or throttling, to a minimum, avoiding recent mistakes of AMD. NVIDIA did slightly bump their temperature target from 80°C to 82°C, which is a very reasonable move and lets the card deliver slightly higher performance without getting as hot as the competition from AMD. Noise levels in idle are quite good, too, but not as low as I would have wished for silent productivity work. Once you fire up a game, noise levels increase, and the card will make its presence felt: It is not whisper quiet or anything close, though its noise levels are very reasonable given the performance. At maximum noise, the card runs slightly below 40 dBA, ending up a bit quieter than the R9 290X in quiet mode and much much quieter than R9 290X in uber mode. Actually, I'm quite positive you could run two GTX 780 Ti cards in SLI without producing more noise than a single R9 290X in uber mode. While the cooler works well, it would be nice to see a backplate giving the product a more complete feel while protecting against damage during handling.
NVIDIA's official MSRP is $700, which is quite high considering you can have the R9 290X for $550 and the R9 290 for just $400. NVIDIA does include a great game bundle with Assassin’s Creed IV: Black Flag, Batman Arkham: Origins, and Splinter Cell: Black List. You could pawn these coupons off to make back $100, bringing the price down some more. It is still a high price, but NVIDIA is simply asking for a price premium for its fastest single-GPU card and lower power/heat/noise, and NVIDIA features like GeForce Experience, ShadowPlay, and G-Sync. Buying the extremely affordable R9 290 for $400 and slapping custom watercooling on it before overclocking it to match the GTX 780 Ti could be an option if you are looking for low noise output, and it would leave you with money to spare. The GTX Titan is largely irrelevant now with its $1000 price. It does offer 6 GB RAM you probably won't need and full speed DP math you most certainly won't need, yet I could imagine some people in the professional/research space to still find a use for it.