NVIDIA GeForce GTX 680 Review: Kepler Debuts - HotHardware

NVIDIA GeForce GTX 680 Review: Kepler Debuts

22 thumbs up

As we’ve mentioned, the GK104 GPU powering the GeForce GTX 680 is based on NVIDIA’s new Kepler architecture. Kepler, however, is not a complete redesign from the ground up. Although much more power efficient and higher performing than Fermi using a number of key metrics, Kepler does borrow heavily from Fermi’s design.


NVIDIA GK104 GPU Block Diagram

The high-level block diagram above shows the overall structure on the GK104. The chip has an arrangement of four Graphics Processing Clusters (GPC), each with two Streaming Multiprocessors, dubbed SMX (a Streaming Multiprocessors in Fermi is called a SM). Within each GPC, there is control logic, plus 192 CUDA cores, for a total of 1536 CUDA cores per GPU. In the previous-gen GTX 580 (Fermi), there were 32 CUDA cores per SM, which were duplicated 16 times within the chip. With the GK104, there are 192 CUDA cores per SMX, which are duplicated 8 times. The structure results in 6x the number of cores per SM(X) and 3x the total number of cores than the GeForce GTX 580.


A Close-Up Of A Single SMX In The GK104 GPU

In terms of its other features, the GK104 has a total of 128 texture units and 32 ROPs. There is 512K of L2 cache on-die, and the GPU interfaces with the GeForce GTX 680’s 2GB of GDDR5 memory over a 256-bit interface. It supports DirectX 11 (not 11.1) and features a PCI Express 3.0 host interface. There are eight geometry units in the chip (Polymorph Engine 2.0) and four raster units (one per GPC). According to NVIDIA, the Polymorph 2.0 engines offer double the primitive and tessellation performance per SM of Fermi.

In addition to having a different GPC and SM arrangement, with Kepler, NVIDIA also minimized the hardware control logic in the chip to bring the transistor count down and Kepler will also operate with a single clock domain—shaders/CUDA cores are not clocked at 2x the frequency of the rest of the chip.

With the GK104, the sum total of all of these changes is a 3.54 billion transistor chip with a die size of about 294 square mm, which is manufactured using TSMC’s 28nm process node. If you’re keeping track, that’s about 770M fewer transistors than AMD’s Tahiti GPU in the Radeon HD 7900 series and a significantly smaller die size (294mm2 vs 365mm2) as well.

Article Index:

1 2 3 Next
0
+ -

I was looking to do an upgrade to this but the price and the fact I need a mobo upgrade (actually whole system) as well pulled me away. I would of done it if I wasn't heading to college next year and I'm going to get a laptop.

0
+ -

Great performance from this card. I have been looking forward to the benchmarks and I am not surprised. Metro 2033 is going to be a tough cookie for some time; but everything looks really solid. Great review Marco :)

+2
+ -

Thanks, man. Killer card for sure.

0
+ -

Yeah the Nvidia prices especially on debut are very high. I was a member at this website last year that did these great give aways pretty often. I was hoping to see them do more give away contest's but they quit I guess as I have not seen any more since last year. The point is I was looking forward to trying to win a system (I could really use one for work and home) with one of these as I cannot really afford it right now. Oh well maybe some other great site will do a contest like that I sure hope so!

0
+ -

Off to peruse this review :)

0
+ -

Great Review Marco, this card is a beast. I have to agree with @LBowen, the benchmarks aren't that surprising and neither is the price. You have to pay to play and Nvidia knows it! I'm happy to see the power consumption and temp results, they are heading in the right direction! No cooking eggs on this card! Looking forward to seeing it in full production and oc'd to the max.

Just out of curiosity I would love to see how this card performed with video and 3d rendering, and yes I understand it's a gaming card not a workstation card.

+1
+ -

Thanks for that hella nice review. Big Green just gave AMD a slap in the face with lower prices and more performance. I can't wait to get my hands on one of these beautiful cards. It will be really interesting to see what the manufacturers come out with for non-reference cards.

0
+ -

.

0
+ -

Yikes...great job NVIDIA. AMD could afford to launch their cards at the exorbitant $599 precisely because NVIDIA hadn't struck yet. Now that Kepler's in town, watch those prices drop.

Gee, AMD can't catch a break, can they? Well it was a fun 3 months or so...

+1
+ -

Looks like the Green machine is back on top,huh? Well I guess they deserve a turn, since it's been awhile.

Let the price wars commence!

1 2 3 Next
Login or Register to Comment
Post a Comment
Username:   Password: