I'm not building an engine, but using one, what i want is to target next gen in assets, so having something that seems like "the best something can buy" today, is usually a good comparison of what average ish will be a few years later,
See but I think it changed since the increase in GPU development has slowed down somewhat. For example, the best in Fall of 2009 was HD5870 CF. It's been 3 years now and HD5870 CF is very high-end for the majority of the PC gaming market outside of these enthusiast forums. The average gaming PC on Steam probably has a GTX560 or HD5770 in it, way slower than HD5870 CF.
Even if you look at the best system from Fall 2010, that's GTX580 SLI. Again 2 years since that point and you'd are looking at 40% faster than that with 670/680/7970 OC SLI/CF.
So your analogy of GTX690 and especially GTX690 SLI representing the mainstream market in 3 years sounds optimistic. The mainstream market performance in 3 years will probably be GTX680/7970 level of performance.
Just to be clear, i'm interested in what could run the best today, not what today's consumer pc look like, so i "do" want to spend time and ressource on 1% or less of the total market, as it may represent a sizeable amount of market in 3+ years so that's definately where i'd like to be for top settings.
Yes, that's another point I don't understand. Why don't you get GTX670 SLI now and then in 6 months reassess the market. I don't understand why you want to buy GTX690/690 SLI and keep it for 3 years for "market estimation purposes". The market is dynamic and as GPUs release, you can sell your old cards and get what you feel is an estimate of the market then. Your idea of estimating the market is wasting $2000 on GPUs that may or may not represent the mainstream GPU performance in 3-4 years? How does that make sense? I don't really understand what you are trying to do. Are you trying to future proof for yourself or seeing how a top of the line system today will fair in 3-4 years? If it's the former, you are wasting $ since future-proofing never really worked. If it's the latter, why not keep upgrading all these years and then see what the mainstream performance is in 3-4 years and then make graphical adjustments accordingly?
If you want to target next generation in assets, you could make those assets as advanced as possible (assuming timeline and budget constraints) since GPUs will only continue to grow and that would make the game look better for even longer (ala Crysis 1). If you want to target the next generation mainstream gamer in 3-4 years, then GTX690 SLI is still too optimistic I feel.
The other difficulty is we do not know how much faster graphics will improve after next generation of consoles launch. For example, if next generation GPUs are 5-10x faster in tessellation and next generation game engines use a lot of tessellation, then may be even GTX690 SLI would be slower than a $200 GPU in 2015. However, if next generation engines are hardly better looking than BF3 is today, then GTX690 SLI would represent something far exceeding the power of a mainstream GPU in 2015 because in that case most people would probably be happy with a GTX680 level of performance. You are basically trying to estimate how fast a $200-250 GPU will be in 3-4 years from now and trying to get a system today that most closely approximates that level of performance? This is very difficult to answer but if look at the last 3 years:
HD5870 ---> HD7970 GE on the AMD side in 2.5 years (that's about 70-75% faster)
GTX480 ---> GTX680 on the NV side in 2.5 years (that's about 50% faster).
^ So you can see it has been very slow progression in the last 2-2.5 years actually.
It's probably expected that GTX880 (Maxwell) will be as fast as GTX690 by 2H 2014 and then let's say GTX1080 is 70-80% faster than Maxwell by 2H 2016. So probably by 2016, a $200-250 GPU will be as fast as a GTX690/GTX670 SLI. But it could be faster depending on what technologies next generation games use that benefit from virtual memory, pre-emption, direct compute, geometry engines of newer GPUs (i.e., they may be exponentially faster with those next generation graphical features).
To show you what I mean HD5870 is exponentially slower in some newer DX11 games but overall it's not 2-3x slower than modern GPUs. That's why your question is so difficult to answer. If in 3-4 years all games use global illumination via DirectCompute lighting effects, then GTX690 SLI would be dog slow even compared to a $200 GPU that's because current GPU architectures are not really built for next generation games 3-4 years from now which would mean an exponential performance hit in the more advanced graphical features such as tessellation, depth of field, global illumination, multiple area lights, etc.
And then you have the other uknown: Will 1080P still be the mainstream resolution, or will it be 2560x1440/1600 or 4K or some other resolution by 2015-2016?
I feel like
GTX670 SLI 4GB sounds like a good balance. It gives you plenty of VRAM to test out your assets with without crippling the GPU (such as would be the case with the GTX690) and should we be wrong about GTX670 SLI 4GB as representative of a mainstream GPU in 3-4 years, you can quickly go out and add a 3rd such card. So it's very flexible in this regard. Going with GTX690 SLI costs more and you don't get the 4GB VRAM flexibility down the line to test your art assets. Also, you still get some benefit of Tri-SLI scaling which is not great but still much better than GTX690 SLI. Plus, you save $700-800 going this path vs. 690 SLI.
You can also use the GPU performance increases in the past to try and approximate where we'll be in 3-4 years (although this is pretty hard but you can ballpark):
Graphics Evolution from 2005 to 2011
Recent Graphics Evolution - 2012
Here is how GTX690 compares to some of those cards
And here is a
GPU database for all the launch dates to give you an idea of how many years it took to get what level of GPU performance increase. In the past it took 2 years to get 70-100% GPU speed increase, but lately this has levelled off.