NVIDA 680 v 670

[Mad Matt]

Nvidia's DP has always been faster than AMD's since the 300 series. By rule, whatever Nvidia can do SP, it can do exactly half that rate at DP which blows AMD out of the water. But like you said, Nvidia cripples it so AMD actually appears to have better DP.

You see, actually I did not know this since before DC I never looked at SP/DP. And never needed any of those cards. So I always thought Nvidia just plain sucks here and is technologically behind. Hey Nvidia, if you are having a PR-bot, take that!

[STE\/E]

**

[pinhodecarlos]

If you see it in terms of energy efficiency (ratio 3dMark/TDP) you have:

GTX 670 is 3.1 % better than GTX 680
GTX 670 is 24.4 % better than GTX 690
GTX 670 is 45.6% better than GTX 590

For calculations I used data from Nvidia page, like TDP and 3DMark DirectX 11 results.

670>680>690>590

In terms of price I don't know.

I don't own any CUDA card capable but when I am looking to buy a CPU I usually look for the most energy efficient one. This means I need to know a ratio of work/power. For me it is not important to have the fastest card or CPU but instead the one that for the same amount of work it uses less energy. Finally I look at the price.

For example, I would buy two GTX 670 instead of one GTX 590.

[DrPop]

Hmmm...interesting way to do it. I think you might get a more accurate result by using the GFLOPS rating for each card, instead of the 3DMark score. I would base my choice on GFLOPS / Watt. That 590 will out crunch the 670 and 680 by a huge margin, because it is a dual GPU card.

[Mad Matt]

If you see it in terms of energy efficiency (ratio 3dMark/TDP) you have:

GTX 670 is 3.1 % better than GTX 680
GTX 670 is 24.4 % better than GTX 690
GTX 670 is 45.6% better than GTX 590

Surprising point of view. So far I thought SP/DP per watt would be the measure. How close is 3dMark to DC performance? AFAIK it's a benchmark not that really far away from it - besides project specifics. E.g. the 560Ti was bang-for-buck king on PG and the 570 on GPUGRID.

[John P. Myers]

Hmmm...interesting way to do it. I think you might get a more accurate result by using the GFLOPS rating for each card, instead of the 3DMark score. I would base my choice on GFLOPS / Watt. That 590 will out crunch the 670 and 680 by a huge margin, because it is a dual GPU card.

Correct. GFLOPS/watt is what's important to crunching. DirectX benchmarks are purely for graphics capabilities and do not give an accurate representation of what they can do in the BOINC world.

GFLOPS/W
690 - 18.74
680 - 15.85
670 - 14.47
590 - 6.82
580 - 6.48
570 - 6.41
480 - 5.38
470 - 5.06

For us, clearly the 690 is far more efficient.

690>680>670>590

[Mad Matt]

JPM, thanks for the confirmation of the 'GFLOP' perspective. Yet I am wondering what difference projects may make on top of their degree of optimization, especially regarding the new 28nm GPUs. Probably there is always a compromise solution.

No clue when projects will learn to use 28nm chips correctly, probably long time after 22nm or 18nm is out...

[John P. Myers]

No clue when projects will learn to use 28nm chips correctly, probably long time after 22nm or 18nm is out...

Probably, since projects still don't even use the advanced instruction sets on CPUs, such as AVX or SSSE 4.2 (or even 4.0)

[Mad Matt]

Probably, since projects still don't even use the advanced instruction sets on CPUs, such as AVX or SSSE 4.2 (or even 4.0)

Some folks in DC act like there was time and power to waste. Interestingly DA is the guy who has most time of them all. Then again, they need time to adapt and most projects except for WCG (yes IBM, you deserve the credits...no matter why you are doing this) are not really well supported economically. Or in other words, they just spend all or large chunks of their idle time and money - just like us.

[pinhodecarlos]

Mad Matt,

Despite we all are here as a hobby we need to have an environment concern, at least a tine one...lol

John P. Myers,

GFLOPS/watt is more accurate than my ratio but we can go even further, there's must be an even accurate way to measure the work done, maybe number of work units done per energy consumed per day. Of course these calculations must be taking in care for the project you will be concentrating on.

Carlos