We knew this day would come, but that doesn't make it any less exciting. After all, we've been waiting since Saturday. Today Nvidia launches the just-announced GeForce GTX 690, which packs two full GK104 Kepler GPUs onto one video card-and what a card it is. (For an in-depth look at the GTX 680, the GK104 GPU, and the Kepler architecture, check out the feature story from our June issue!)
Two. Two GPUs.
The GTX 690 is 11 inches long-big for an Nvidia card, but still smaller than the 12.2-inch high water mark established by the AMD Radeon 5970 a few generations ago. As you'd expect, the GTX 690 contains two of the same GPU found in the GTX 680, with a slightly lower base clock-915MHz with a boost clock of 985MHz, compared to the 1,006MHz base and 1,058MHz boost clock for a reference GTX 680. Nvidia says they've built in substantial room for overclocking, too, saying that you can get over 1,100MHz clocks from the stock cooler.
Aside from the slightly lower clocks, the rest of the board's specs are exactly what you'd expect from a true dual-680 configuration: 3,072 CUDA cores, 16 SMX units, 256 texture units, and 64 ROPs. Each GK104 GPU has 2GB of GDDR5 with four 64-bit memory channels per GPU, for a total of 4GB GDDR5 frame buffer for the whole card. The card takes two 8-pin power connectors and has a TDP of 300W.
The GTX 690's reference shroud looks amazing. It's all magnesium-alloy and polycarbonate, with nary a bit of plastic to be seen. In typical Nvidia dual-GPU card design, the intake fan is in the center of the PCB and blows air toward both the front and back of the card-keeping cool air flowing over both GPUs but potentially wreaking havoc on standard front-to-back case airflow.
Here's how the reference design of the GTX 690 stacks up to a reference GTX 680 and an XFX Radeon HD 7970 Black Edition.
Nvidia made sure to tell us that unlike the GTX 590, which was a fair bit slower than dual GTX 580s, the GTX 690 is almost as fast as dual GTX 680s. The card has a TDP (thermal design power) of 300W, so there's plenty of overhead for overclocking.
We haven't had our GTX 690 reference card for very long, but we were able to run our graphics benchmark suites to see how it stacks up to two other dual-GPU configurations: stock-clocked GTX 680s in SLI and stock-clocked Radeon HD 7970s in CrossFireX. We tested all three configurations in the same test setup, which consists of a stock-clocked Core i7-3960X on an Asus P9X79 Deluxe board with 16GB DDR3/1600 in a Cooler Master Cosmos II chassis with a 1050W Thermaltake PSU and a 256GB Samsung 830 Series boot SSD. We used the latest Catalyst drivers from AMD for the 7970s, the latest WHQL drivers from Nvidia for the 680s, and the pre-release drivers Nvidia provided for the GTX 690.
Running the Numbers
True to Nvidia's claims, the GTX 690 at stock clocks was only a few frames per second slower than dual stock-clocked GTX 680s. The greatest difference we saw was 4fps in Just Cause 2's Concrete Jungle benchmarks, and the GTX 690 was actually very slightly faster than the dual 680s in Metro 2033.
How do the Nvidia SLI solutions compare to two reference Radeon HD 7970s in CrossFireX? They trade blows: the Nvidia cards are faster in both 3DMark Performance tests, Unigen Heaven, Far Cry 2 (barely), and Batman: Arkham City, while the Radeons win in Shogun 2, STALKER: Call of Pripyat, Just Cause 2, and Metro 2033.
Twice as Nice for Twice the Price
We haven't even started exploring the GTX 690's overclocking options, nor have we done all the benchmarking we want to do, but from our initial tests it's an astonishingly fast single-card SLI setup that takes up less room and draws less power at idle than multi-card setups. And, of course, if you have two of them you can enable Quad SLI. Yes please.
For the price of two GTX 680s, you get the power of two GTX 680s on one board. That seems pretty fair. The GTX 690 is available starting today from a slew of Nvidia partners with an MSRP of $999.