Maxwell 2
21 Sep 2014 9:48:22 UTC
Topic 197729
(moderation:
Hi guys,
is anyone already running a GTX970 or GTX980 here? They should provide amazing performance with excellent power efficiency.. but it would surely be good to check real world results :D
MrS
Scanning for our furry friends since Jan 2002
Maxwell 2
)
Gee, I think we are about two days after product launch, with both NewEgg and Amazon showing near zero in-stock availability and a near complete absence of user comments. So anybody posting results here either had a preview card or is very fast on their feet.
But I am very, very interested myself. My first generation Maxwell (GTX 750, not the "i" model) purrs along with useful output at very low power in a machine which was constrained by use and location to be low power. I'd be interested in the possibility of swapping in a 970 or 980 for the 660s on my two primary crunchers if the Einstein performance turns out to be good.
In Germany they are
)
In Germany they are available, but not in the shops wiith the lowest prices. That's why I can "afford" the time to wait a little bit longer and see how good they really are.
One person at GPU-Grid already got one, but the app seems to need an update.
MrS
Scanning for our furry friends since Jan 2002
To me the Maxwell promise of
)
To me the Maxwell promise of greatest interest is low power consumption in respect to performance.
Amid the generally glowing 970/980 reviews I have read since I was alerted to the product release by this thread, the one discouraging word was the position taken by a Tom's Hardware review that much, or perhaps even all, of the 980 power consumption advantage in certain comparisons came from superior management of and adaptation to rapidly changing power requirements. The reviewer specifically predicted that work loads with steadier high computation requirements, such as CUDA work, might see much less improvement than more game-like rapidly changing work.
I don't have much of a clue as to whether an Einstein Perseus work load is stable or rapidly fluctuating at the granularity relevant to this consideration, so don't have a preconception on what if anything this claim means for Einstein, but it does increase my interest in early reports of Einstein performance of the 970/980 products. I'd be especially pleased if early reporters would report power consumption observations. It would also be helpful if the specific model and rating of power supply used were mentioned.
I imagine our German contributors who are paying their own residential power bills have a yet stronger reason than I for power consumption interest. After a recent rate rise my marginal cost for power if I trip over the line into the next rate tier is about 16.5 cents US/kWHr. So the roughly 100 watt savings claimed for the 980 vs. the 758i is worth between $100 and $150/year to me. So even on a multi-hundred dollar US card the power cost is worthy of some consideration.
I'll be away from home or distracted by major household events for much of the next two months, but unless I learn something bad about this I expect to install and report on one card by mid-November. I hope others will be before me.
RE: So the roughly 100
)
Where I typed "758i" I intended "780 Ti"
http://einstein.phys.uwm.edu/
)
http://einsteinathome.org/host/11669559/tasks
Not my unfortunatelly.
Link taken from GPUGrid forum.
Card is 970.
RE: http://einstein.phys.uw
)
That user is a bit curious--they have barely used the 970 here at Einstein--just four tasks, all returned already and none in process. That user shows only one other recently active host on Einstein, with just six tasks, with an AMD Tahiti for a task reported on September 19, but a GTX 750 Ti for one reported on September 16. They seem to be dipping their toes in the pond.
That user has processed much more 970 work on SETI--maybe on other projects I did not review.
Can anyone here spot some useful conclusions from the reported work? I guess the first order conclusion is that one can get validated Perseus results, and can return apparently successful completions on BRP4G with a currently available driver for the 970, and current Einstein software.
I was just about to post
)
I was just about to post those result links :)
Taking useful data away from it is difficult, though, as I'm not usually running Einstein GPU work on anything else than Intel GPUs. I tried to find comparable hosts (nVidia GPU, Win) and first looked at BRP Persus results:
GTX970: 5600 s/WU
GTX780: 21000 s/WU
GTX680: 15000 s/WU
While I want MAxwell to be fast, I don't think this is the whole picture. Since the user reporting values for the GTX970 seems to be at home at SETI and new to Einstein, we can assume he was running 1 WU at a time. But how many did the others run? 2 on the GTX680 and 3 on the GTX780?
Comparing Arecibo tasks obviously poses the same challenge:
GTX970: 1880 s/WU
GTX780: 4500 s/WU
GTX680: 3300 s/WU
... which could again match the number of concurrent WUs speculated previously. Note: I just averaged most results by eye, don't use them as precise numbers.
Can anyone extract better information out of these results?
MrS
Scanning for our furry friends since Jan 2002
From GPUGrid
)
From GPUGrid forum:
It's for one WU at once.
Now two are running for that host.
Card looks nice, but still it's below halve my AMD 7950@1100 here at Einstein.
Here is a second set of data
)
Here is a second set of data for the new GTX 970.
Platform is i7-4770K running at 3.9GHz with DDR3-1600 running no CPU jobs for this set of tests. OS is Windows 7. Idle power draw is 61 watts measured at the wall with a kill-a-watt meter. Power supply is a Seasonic platinum. GPU power usage was difference between GPU running and idle.
All tests run with the new 344.11 driver on both 750 Ti and 970. Right now I'm running 2 jobs at a time which was a good fit for the 750 Ti that I will compare the 970 against. I'll switch over to 3 jobs at a time on the 970 tonight. Using BRP5 for current tests. The GTX 970 is MSI's high end factory overclocked version. MSI Afterburner version 3.0.1 was used to measure GPU temps and usage.
750 Ti = 14,922 secs; GPU Usage = 97%; temp = 46 C; watts = 44
970 ... = 8,677 secs; GPU Usage = 95%; temp = 61 C; watts = 123
750 TI Daily Credits = 3333 * 86,400 / (14,922 / 2) = 38,597
970 ... Daily Credits = 3333 * 86,400 / (8,677 / 2) = 66,376
750 Ti performance / watt = 38,597 / 44 = 877
970 ... performance / watt = 66,376 / 123 = 540
I was hoping to see the 970 scale up to the same performance / watt as the 750 Ti, but as you can see above that did not happen. But just to throw in a comparison against my current AMD 7970, here is that comparison.
7970 performance / watt = 95,024 / 175 = 543
So the new 970 is comparable on performance / watt to the 7970.
Robert, That is a very
)
Robert,
That is a very nice comparison. The GTX 970 should be a killer on Folding, but more ordinary on the other projects. That is the same conclusion I have reached for my GTX 750 Ti's, having tried them on Folding, E@H and GPUGrid.