monsta :
In the reviews and benches the 680 with 2gb of vram still beats the 7970 with 3gb in many games, it only trails the 7970 by less than 2 fps in Metro @ 5760x1080, not a significant difference.
http://www.tomshardware.com/reviews/geforce-gtx-680-sli-overclock-surround,3162-10.html
It shows weakness is what I said, not that it's no good right now. Also, at 5760x1080, it's a 5FPS difference (not a small difference when it's 40FPS compared to 35FPS) if you use AA. Unless you don't like AA, there's no reason to not include that in the measurements and all the reason to include it. I also said that it would probably take a few years for the 680's memory problems to be a serious problem. Problems it has right now is that it can't handle AA as well as the 7970 in some games at some resolutions and quality settings in some games.
The 680 has a VRAM capacity bottleneck (it's VRAM bandwidth is a little low for it's GPU performance, but it's acceptable and it can't make performance drop like a rock into unplayable like a capacity problem can). Some games, such as Metro 2033 and BF3, can push past 1GB of VRAM usage at 1080p. Going to higher resolutions does not increase VRAM usage linearly from this point (or else the 680 would be screwed so bad that Nvidia wouldn't have even released a 2GB version and would have only done 4GB), so the 680 is okay at high resolutions in some games if you don't overload the AA and quality settings. Hard OCP did a test on 5760x1080 and had to lower AA in some games (even done to just FXAA or no AA some of the time) because the 680s don't have enough VRAM.
The 680 clearly has a slightly more powerful GPU for 32 bit math (gaming right now is almost purely 32 bit math). However, the 7970 has almost 6 times more 64 bit math performance than the 680 (not a joke, it's a little over 5.5 times more). Games are starting to incorporate more 64 bit math than ever, although that trend has not come close to peaking yet. Civ 5 is a start of it, but it only has a little (not so much that the 680 can't handle it). Even the 580 has more 64 bit math performance than the 680 (almost double).
Kepler, in it's current form, is best skipped if you want your computer to last longer at maxed out settings in games without an upgrade. The only two reasons that the 680 beats the 7970 in Civ 5 is because Civ 5 doesn't use a whole lot of 64 bit math and because the 680's 64 bit math is done on cores separate from the regular, 32 bit cores (Kepler cores in the consumer cards can only do 32 bit math. They are optimized for it, hence their advantage in 32 bit math over GCN and all other GPU architectures). The 64 bit cores are different hardware, so using them does not slow down 32 bit math much at all (it only slows it down a little because of the 680's low memory bandwidth).
The next generation of Kepler might have more 64 bit cores per card, alleviating this problem, and more memory capacity and bandwidth. However, in it's current iteration, it is not the best way to go, in my opinion. The 670 is a better choice than the 680 solely because it is slower, but has the same memory capacity, so it's capacity will not be as much of a bottleneck. However, it still has the low 64 bit math problem. In fact, even the 7850 and 7870 beat the 680 in 64 bit math (although due to them, like Kepler, using non-compute oriented GPUs, they don't come close to the 7950 and 7970).
The Kepler Quadro and Tesla cards will have incredible 64 bit math, but Nvidia decided that they don't want consumers to have it because they want you to pay thousands of dollars for overpriced professional hardware to get it.
Next time, you might want to make sure that you're not arguing with someone who has some computer engineering knowledge when you want to argue about computer hardware.
Facebook
Twitter
Instagram