|NVIDIA GeForce GTX 680 Graphics Performance|
|Reviews - Featured Reviews: Video Cards|
|Written by Olin Coles|
|Thursday, 22 March 2012|
Page 3 of 15
NVIDIA Kepler GPU Details
Like any high-performance machine, it's what hides under the hood that counts. On the outside, NVIDIA's GeForce GTX 680 is merely another video card that shares a similar profile with almost every other product released over the past four years. But on the inside, NVIDIA's codename "Kepler" GPU architecture reshapes the internal landscape. Once the plastic shroud is removed, you'll notice a large heatsink covering half the PCB while a 60mm blower motor fan is positioned nearby. Beneath the cooling equipment is a 12-layer printed circuit board (PCB) to ensure the highest signal integrity, and to help disperse heat more effectively across the PCB.
Made popular on high-performance CPU coolers, NVIDIA uses embedded heat-pipe technology for GTX 680s thermal management system. Kepler's lower TDP already reduces heat output, so using the larger and more expensive hollow vapor chamber design found on GTX 580 became unnecessary. The thermal management system on GeForce GTX 680 actually falls somewhere between that of the GeForce GTX 560 Ti and the GTX 570.
With the heatsink removed, NVIDIA's GK104 28nm processor is exposed. Packed with 1536 CUDA cores across eight SMX units, the GK104 GPU is assigned a base clock frequency of 1006 MHz which boosts to 1058 MHz when needed. Kepler's architecture is comprised of 192 CUDA cores, 16 texture units, and polymorph engine (2.0) per SMX cluster. By comparison, the increase in cores balanced by a reduction in core speed helps Kepler achieve far more efficient (2x according to NVIDIA) performance per watt when compared to the Fermi architecture.
The memory subsystem has been tweaked on GeForce GTX 680, allowing the 2048MB GDDR5 video frame buffer to produce 192.26 GB/s total memory bandwidth at an impressive 6008 MHz data rate. Four memory controllers combine eight GDDR5 ICs for a 256-bit memory lane, which moves at record speeds but still operates more efficiently than previous designs to yield a fill rate of 128.8 GigaTexels per second.
Examining the printed circuit board (PCB) reveals a few new changes, namely the absence of an aluminum plate heatsink cooler and the inclusion of a Richtek Technology Corporation RT8802A multi-phase synchronous PWM advanced digital power controller with over-volting capability. We examine power consumption later on in this article, using 3DMark11 to represent real-world loads.
In the next section, we detail our test methodology and give specifications for all of the benchmarks and equipment used in our testing process...