Today Nvidia introduced the first card in its new 7 series of graphics cards, the GeForce GTX 780. Based on the name you probably expect an entirely new architecture, but the GTX 780 is based on the same GK110 GPU that's the foundation of the GeForce Titan. The new card has slightly fewer CUDA Cores (Nvidia's term for its unified shaders), somewhat less video memory and lower clock speeds. While it's not the 1,000 euro investment of the Titan, it still carries a hefty pricetag. Today we will take a closer look at Nvidia's new card and find out what it's capable of.
The GeForce GTX 780 doesn't have that many surprises in store in terms of hardware. Like the GTX 680 and the Titan, along with the other 6 series graphics cards, it's based on Kepler architecture.
If you're interested in all the details of the Kepler architecture and the GK104 and GK110 GPUs, please read our reviews of the GeForce GTX 680 and the GeForce GTX Titan. This review will focus solely on the GTX 780, but we will also compared it to the GTX 680 and the Titan on certain points.
Like the Titan, the GTX 780 consists of 7.1 billion transistors. That suggests that the card either is a GK110 with a few inactive cores or, more likely, a GK110 that didn't make it through Titan certification but still had enough working transistors to be used in a lower positioned card. Where the Titan uses 14 out of the 15 operational 'Streaming Multiprocessors' or SMX units, with 192 Cuda cores each, the GTX 780 uses 12. That equals 2,304 Cuda cores or shader units for the GTX 780. That's a bit less than was speculated earlier, and exactly 50 percent more than the 1,536 shader units in the GTX 680. Nvidia promises a performance gain of 70 percent compared to the GTX 580 and 34 percent compared to the GTX 680.
The Cuda cores in the GTX 780 run at 863 MHz, which can increase to 900 MHz when the temperature allows. This is made possible by Nvidia's Boost 2.0 technology, used in the Titan already. The Titan was slightly slower, with 837 MHz and a boost of 876 MHz. The GDDR5 memory is equally fast as on the Titan, at 1,502 MHz. This translates to 6,008 MHz since four instructions per clock can be processed. The GTX 780 has half the amount of memory of the Titan, but it's still an impressive 3 GB using a 384-bit memory bus. The available memory bandwidth is the same as Titan's, 288.4 GB/s.
The TDP of the GTX 780 is also the same as Titan's with 250 watts. Nvidia mentions in the fineprint that this is not the maximum usage, and that it can turn out to be higher in synthetic benchmarks such as Furmark. The GK110 automatically clocks itself down when the temperature exceeds 95 degrees Celsius.
The recommended price for the GTX 780 is 539 euros, excl VAT, which amounts to about 650 euros. That's more than was originally expected, initially it was rumored to be between 499 and 599 euros. The number of Cuda cores is lower than initially assumed as well. It's a lot cheaper than the Titan that costs about 1,000 euros, but of course also a lot more than an AMD Radeon HD 7970, which costs between 375 and 425 euros.