you rather get the 192 bit version and OC the vram to what ever you need to match the bandwidth of gtx 670, (i say you needed 6 ghz but i was wrong, you need more, i dont know how to calculate). Cutting down cuda cores requires heavier OC to reach the 670 gtx
Clock to clock, the SMX disabled on the 670 just bring 2% maximum of difference in performance, hence why the 670 Oc retails models ( Asus DirectCUII ) are faster of the 680 by a good margin ( and even faster of some OC 680 ) ... It is not a problem for a 660TI with One SMX disabled and the exact same amount of Vram, the same memory controller to go beat the 670.. Ofc its needed to see how TMU, etc will act with 2SMX disabled. The 670 was normally set with a Turbo clock reference of 980mhz, but in reality the card work mostly higher of 1084-1100+mhz... if this card is allready set at 1123mhz Turbo ( at least on the model they describe ), just put ~1180mhz and the card is allready faster of the 670.
More things in my experience are shader limited rather than memory bandwidth limited. Only in the most extreme cases with low end cards I've ever seen memory bandwidth be the limiting factor. That's too large of a hit to CUDA cores just to get a memory bus 64 bits larger, I think the benchmarks will show as much. Not to mention with almost ever card I've had the memory OCs more than the core and shaders.
going from the 680 to 670 in shader count had almost no change in performance like 2-4%. 192 bus imo is going to gimp a card more than the shader count difference. the 256bit bus on a 680/670 has only the same memory bandwidth as a 580 (384 bit, 4ghz clock speed)
hmm 256 bit bus... If i cant resolve this 670gtx shrieking like banshee under any kind load, by Rma my PSU or the card it self I will just get refund on the 670 and get 660ti with 256bus if i paying 400$ gpu i expect the damn thing not to have caps/chokes that shriek like banshee.
Is the 660 192 bit version that much slower than the AMD 7870, that it would not benefit from having the 256 bit clocks and power connectors. 915/980 vs 980/1123 clocks.
670 review date now, so find the good one where they have test it will be a bit hard, but well i just find this.. The Evga is working at 1046mhz turbo (well surely a little bit higher, as we know all turbo is not fixed and the number given is allways the min. i will bet for a 1084-1100mhz ), so maybe a bit slower of the 680 or equal. I will not even take the time to calculate the % ratio, it is obvious looking the numbers. (outside Dirt3, where they have not apply AA the difference max is around 2fps each time ) I know it is really hard to get the real clockspeed with how Nvidia have set the turbo, and particulary on press samples, but this is just for give an idea.
8% less performance than a 670 is what, 10-15% behind a 680? For 200 less dollars? So with OCing were talking 680 performance (or pretty close to it) while paying 40% less? Sounds good to me, now lets hope it's true.
its widely known a 670 is just a couple % slower here and there than a 680 http://www.anandtech.com/bench/Product/555?vs=598
If Tri-SLI scales perfectly, GTX660Ti Tri-SLI should have similar or better performance compared to GTX680 SLI unless GTX660Ti is 50% slower than GTX680. If it's only 20-25% slower....Tri-SLI GTX660Ti would be faster....but the cost would still make it questionable compared to GTX680 SLI.
Yes I took a look at some benchmarks, the average increase in games seemed to be at least ~10%, certainly a lot better than 2-4%. Of course I know the point you're trying to make is that the difference is small. The article also makes it sound like the memory bandwidth was too much of a hit. But I always keep the maximum OC in mind, I wonder how that would have turned out, but we won't find out. I doubt nVidia will release two versions.
http://www.anandtech.com/show/5818/nvidia-geforce-gtx-670-review-feat-evga/5 these benches are all over the net, the difference is very little
Why are you cherry picking things as if you're AMD or something, that very link shows some with a 20% difference. No one is disagreeing that the average difference is very little, just the 2-4% number is far from an average. Edited mistake.