why does 660 Ti performs as 560 Ti and not faster?
7 Nov 2012 10:02:38 UTC
Topic 196613
(moderation:
Good day,
i have discovered next problem:
average WU runtime on my 560 Ti is 1100 sec for single WU and 2100 sec for 2 simultaneous WUs.
660 Ti according to wiki has almost double GFLOPs performance than 550Ti (2400 vs 1280) but unexpectable has the same runtime for 1 and 2 WUs?
why? cuda app v 1.28 has no optimization for the new fermi GPUs?
should i return my new 660 Ti to the shop and stick to rather old 560 Ti?
why does 660 Ti performs as 560 Ti and not faster?
)
Hello Astrocab,
yes, this is a question that was asked earlier sometimes. The anser I got is: 5xx series cards use double clock-rate for the core, this is why they are much faster with the same amount of cuda cores.
Alexander
RE: Good day, should i
)
Yes, and I'd recommend 560 Ti 448 core. Of course, this is not truly 560 but the RAC is good. Look at the statistics on the main page for the best computers.
560ti has 1645 MHz clock for
)
560ti has 1645 MHz clock for shaders, 660to has only 915 MHz, i.e. almost twice slower,
but(!) 660ti has 1344 shaders and 550ti has only 384, i.e. 660 has 3.5 times more.
and (1344 * 915 MHz) / (384 * 1645 MHz) = 1.95
technically 660ti is 1.95 times faster than 550ti
but in our application didn't see this
i think problem is in application optimization.
somebody know who is developer?
As I can see E@H application
)
As I can see E@H application is not using all available CUDA cores because I can play some not very intensive games such as Fallout NV along with 2 WUs on the same 660 Ti. It was almost impossible on 560Ti with the same amount of video memory.
ok, i think main question
)
ok, i think main question should be directed to app developers. many of us need kepler-optimized app which will utilize full potential of new kepler gpu
BRP4 has never made good use
)
BRP4 has never made good use of high end GPUs. The problem is that it's not a pure GPU app and as the GPU gets faster communications with, and computations on, the CPU become an increasingly large fraction of the runtime.
my CPU is loaded about 20%
)
my CPU is loaded about 20% while crunching and it is cool. so does 660ti.
and i want to make them work at full potential
Here you
)
Here you go;
http://stackoverflow.com/questions/10765518/how-to-quantify-the-processing-tradeoffs-of-cuda-devices-for-c-kernels
Kepler is designed to be a gaming platform first, and a CUDA device second.
HOWEVER, to move things forward, there is going to be divergence, I suspect in the same way with FireGL/Quadro cards vs. Geforce/Radeon.
The design of Kepler vs. Fermi is more the issue than the card specifications are. I am running a couple 550Ti and a couple 660Ti Superclocked.
I can tell you the 550ti runs a single WU in 28 mins. The 660ti SC in about 25-26 minutes. In other words, the $100 card is nipping on the heels of the $300 when it comes to crunching. Side by side comparisons in Torchlight 2 however show the 660 having more than double the minimum frame rates of the 550Ti.
The specs on the 660ti are way beyond the 550 in speed, fill rate, having 6x the cuda cores and so forth. The only thing close is memory bandwidth at 192 bit. Even if you follow the explanation given in the link above, it basically states that Kepler running at half the cores would still mean that the 660ti is making use of 3x the cuda cores of the fermi. So this isn't the issue either.
Consider another comparison; my GTX 260 maxcore 55nm vs 550ti. Again, in gaming the 260 outperforms the 550ti (only by a little, and only on some titles). But the 550ti is double the crunching performance. Why bring this up about an older card? Because performance is relative to what you're doing, rather than the raw specs of the card. Neither card specs, cuda cores, nor price is any indication of how well a card will crunch WU vs another card.
I built my main machine as a gaming platform first. I went with 2x 660Ti in SLI mode. I let Einstein run on it when I'm not gaming so whatever it does, is fine by me. I discovered the 550Ti by accident when replacing a bad card in a second box. The crunching performance was great. If you want to buy cards specifically for crunching, the 550ti is probably the best bang for the buck. Anything more and you're spending a ton of money for a tiny upgrade.
All in all, I don't think that the Kepler is being underutilized, or that the app is specifically designed for Fermi. Fermi is just pound for pound a better cruncher than Kepler. But when it comes to games, Kepler dominates Fermi cards. As previously mentioned, I expect to see a divergence based on use as was the case with workstation graphics cards.
but 550ti has only half 560ti
)
but 550ti has only half 560ti performance at little less price. so, 560ti is more preferable here.
and it is seemed fermi card now are the best solution to run E@H?
A 560Ti is about 33% faster
)
A 560Ti is about 33% faster at DOUBLE the cost. Speaking pure bang for buck the 550ti is better. You are spending 2x as much for a 1/3 gain.
For crunching all day the 560ti is better. But where do you draw the line? Get a quad setup of GTX 580s and be done with it?
It comes down to what you're trying to do, what your goals are, and what your budget is. As I stated before, you have to spend a lot more money to get only a little gain. And, the higher you go the worse it gets.