GTX660 Ti on the way

[Snow Crash]

I have one on the way and should be able to post some results next week.
For a start I'll be plugging it in to the same system as a GTX670 where both will be running at PCIE2 X8 so we can get a decent baseline.

- PCIE3 systems are kind of hit and miss right now as while they may be electrically capable most are not running at PCIE3 without a reg hack that may or may not work on a model by model basis.

[shadowwind]

cant wait to see some stats from this card i hope you run the 660 by itself to see what power it consumes on its own off the wall.

[Snow Crash]

I'm a bit to tired and busy to dig through and compile a big 'ol pile of stats but what I can say is that is runs a bit faster than a GTX480 at about 1/2 the power draw.
I have it running GPU@ 1215, VRAM@ 3206 and it is rock solid stable.
Comparisons done on NATE WUs as they arer very consistent in size and runtimes ...
660Ti seem a bit hampered by the 192bit bus ... MCU is at 41% compared to 24% and the GPU Utilization % is at 92 as c0mpared to 95.
While they are both running at 8X PCIE2 so the comparisons are relevant, I'm sure it would do better (and even more so compared to 480) if you run it on x16 by itself or even better if you could go full bore PCIE3 @16X

[Snow Crash]

I've pushed it a bit higher ... still stable with GPU @1270 and MEM @3225.
Can't go much farther as I'm hitting ~108% power :-)

<edit> just noticed after the last reboot I forgot to push the fan manually, it is running 50% w/ temps at 64c <edit>

This is making me think I might be able to push then 670 a bit more but it has been rock solid for so long with GPU @1250 and MEM @3200 that I'll be pissed if I start crashes just for the sake of 20hz more.

[werdwerdus]

if it ain't broke, you haven't fixed it enough yet

[werdwerdus]

Would you more recommend the 660 or the 670 right now? I will be watching the deal sites for either one

[werdwerdus]

well on Amazon the EVGA 670 has a $20 coupon you can "clip". Comes to $359.99 with free shipping. Should I bite or is this not a very good deal?

[Snow Crash]

I feel the bang for the buck is with the 660Ti but 670 are also very good compared to the 680. I have one of each and if I was to get another one right now I would probably lean to the 660 to save a few bucks both on the purchase and also going forward on my electric bill.

I took a quick look and it appears that 670 is a pretty good deal but comes with a stock cooler and I'm not sure it is going to be all that much better than a 660 anyway

Not that stock coolers are bad but I decided after going reference on the 295, 285, and 480 that custom coolers from reputable companies perform so much better and seeing as how keeping the GPUs cool is even more important (frickin drivers forcing clocks down when the GPU gets warm crap) I'm not likely to ever go back.

[Otis11]

Looking at the specifications it seems the only difference is memory bandwidth? Are the core clusters the same size/ratio? If so I'd say go for the 660 every time!

[jimba86]

I waited results,should be great to see what this card can do!

[Snow Crash]

Looking at the specifications it seems the only difference is memory bandwidth? Are the core clusters the same size/ratio? If so I'd say go for the 660 every time!

Shader count is the same and while mem bandwidth is down to 192 and ROPs are down to 24 I don't think either of those matter anymore for GPUGrid.

I waited results,should be great to see what this card can do!

While not quite an apples to apples comparison ...

NATE WUs are pretty consistant in size and runtime ...

670 is running on an i7-920 at 4.0 GHz PCIE2 x16 = 14,000
660 is running on an i7-980x at 3.8 GHz PCIE2 x8 = 15,750

Both are on Win7x64 and I'm confidant that the difference is in the CPU speed and the PCIE2 x8. There are a few other minor differences ... RAM on the 980X is slightly looser timings and QPI is 1.5x RAM instead of 2x RAM as on the 920). And no, I don't feel like swapping cards around as I'd have to lug the monitor and cards between my temp study in the basement (summertime heat) and my real study (2nd floor) and then back again ... I'm just too lazy to get "real"

[werdwerdus]

looks like the only difference from the 660 to 670 is the memory bus width and the number of ROPs. I will definitely be getting a 660 next!

[Snow Crash]

I'm going to try and fire up 2 instances of PAOLA WUs later today on the 660Ti. To get that working as best possible I should pull the other card out of the rig, switch slots, update driver, etc. (that's why I say try ... my wife might have other plans for me wwhen I get home ) If it all works out I will start deleting NATE's and run PAOLA exclusively to see if we can clear this queue.

side note: The card I'm pulling is a GTX480 and I'm thinking about decomissioning it, aPM me if interested.

[werdwerdus]

thanks, I got stuck with multiple PAOLA wus yesterday, my points really tanked! Any idea how many WUs are left on her project?

[Otis11]

I'm going to try and fire up 2 instances of PAOLA WUs later today on the 660Ti. To get that working as best possible I should pull the other card out of the rig, switch slots, update driver, etc. (that's why I say try ... my wife might have other plans for me wwhen I get home ) If it all works out I will start deleting NATE's and run PAOLA exclusively to see if we can clear this queue.

side note: The card I'm pulling is a GTX480 and I'm thinking about decomissioning it, aPM me if interested.

How are you going to fire up 2 WUs on a single card?

And good luck with that... clearing us of the slow ones will only help us all get better utilization!

[werdwerdus]

http://www.gpugrid.net/forum_thread.php?id=3118

[Otis11]

http://www.gpugrid.net/forum_thread.php?id=3118

Ah! So from reading through it, it looks like there's no way to have the computer automatically switch for the PAOLA WUs... That sucks... lots of manual overhead...

[werdwerdus]

yep that is why I won't consider it. I work 9-10 hours days and driving is almost 1 hour round trip so I'm away from my systems a lot. plus with the "normal" workunits, running 2 on one card cuts gpu usage to about 40% and crazy unresponsiveness etc

[Snow Crash]

I did manage to get things working last night and while I can say that GPU utilization went up but I think the best this is going to help is that the card no longer slows down on it's own which happens when nvidia decides it's not busy enough to warrant running at the clicks I set
Normally a GPUGrid WU (any kind) uses a full CPU thread if you have a 6XX card and Win7 (not sure on other OSes) but when I'm running 2 at a time it splits it's CPU usage between the 2 !!! argggghhhh ... these WUs are slow precisely because they have computations that need to happen on the CPU and (yes, GPUGrid posted that info but we all just want to biotch about them running slow). I have tried repeatedly monkeying with the avg cpu and max cpu tag values in the app_info but have been unable to convince the app to take more CPU. Even turning off my CPU tasks the ACEMD is still only taking .5 thread a piece.

While I have a day job and a 2 hour roundtrip commute these WUs run for a very long time and I should be able to do the manual, manual , fetch, abort, fetch between late night and early mornings. Now I'm trying to figure out if I am going to take the same hit and do thi on my 670 also

[werdwerdus]

maybe you should try SWAN_SYNC 0, it may not do anything but it could work

[jimba86]

would the lower bandwith of the 660 effect it at all like what happened with 460 768?

[Snow Crash]

maybe you should try SWAN_SYNC 0, it may not do anything but it could work

Good thought ... I'm ready to try anything. Running 2 at a time is going to take 30 hours which kills me even more on points but is actually faster than the 670 that is going to take 16 hours for the one it is running right now. Depending on how well I do first thing in the morning at grabbing 2 more PAOLA's I will do this on the 670 also. Hell I'm even thinking of dropping the 480 back in because I will be able to run it lower voltage.

would the lower bandwith of the 660 effect it at all like what happened with 460 768?

I don't think that is a factor for crunching GPUGrid but now I'm tempted to turn the 660 back to only 1 WU at a time and run 1 NATE (they are very consistant in runtime) to see how it really does against the 670 as they now are both running on PCIE2 x16 w/ CPU at 4 GHz.

I wish I had more time and another rig (don't temp me) so I could try out some linux ... now I'm thinking too much, I have a CPU, PSU, HSF, all I need is an x58 MOBO and some ram ... no no no, save it for some smoking haswelll you jaskasci

[Snow Crash]

LOL ... I take on too many different things at the same time

I did run a NATE on the 660 but had a number of *issues* with my rig so I had a couple of restarts while running it which only serves to lengthen the runtime as the WU had to pick up from a checkpoint more than once

I've now started running double PAOLAs on my i7-920 Win7 x64 in conjunction with CPU tasks from other mulitple other projects ... I left 1 more core free than necessary and something odd is happening ... instead of splitting the core in half for the 2 instances of ACEMD it is using a bit more than 1/2 for each and I've already finished 2 in 14 hours and am on the same trajectory for the next 2!!!

If I was to start wildly seculating as to why 1 with no CPU tasks takes 16 hours and 2 with some CPU tasks takes 14 hours each, besides that the GPU does not slow down, it would be that because these WUs require actual calculations on the CPU and perhaps the CPU is trying to be smart and use cache that is stale and possible causing mispredictions where if I am running 6 low weight CPU tasks (not 7 threads WCG Schistosoma) that it knows the cache will not help and reloads the calcs entirely so it doesn't have to pay for the cache miss and the misprediction cycles???

Otis, you know way more about the real guts of a CPU ... what say you?

side note: manual intervention to get double PAOLAs is a royal pita

[werdwerdus]

So I was a little out of my mind last night... won't go into detail but let's just say I had a "good time"

Woke up, checking my email, see heading of "Your Order with Amazon.com". Hmm don't remember ordering anything this weekend. Open up the email... and



I suppose I could cancel it seeing as it's still Sunday, but what the hell, right? The science gets done, we make a neat gun, for the people who are still alive.

[Otis11]

LOL ... I take on too many different things at the same time

I did run a NATE on the 660 but had a number of *issues* with my rig so I had a couple of restarts while running it which only serves to lengthen the runtime as the WU had to pick up from a checkpoint more than once

I've now started running double PAOLAs on my i7-920 Win7 x64 in conjunction with CPU tasks from other mulitple other projects ... I left 1 more core free than necessary and something odd is happening ... instead of splitting the core in half for the 2 instances of ACEMD it is using a bit more than 1/2 for each and I've already finished 2 in 14 hours and am on the same trajectory for the next 2!!!

If I was to start wildly seculating as to why 1 with no CPU tasks takes 16 hours and 2 with some CPU tasks takes 14 hours each, besides that the GPU does not slow down, it would be that because these WUs require actual calculations on the CPU and perhaps the CPU is trying to be smart and use cache that is stale and possible causing mispredictions where if I am running 6 low weight CPU tasks (not 7 threads WCG Schistosoma) that it knows the cache will not help and reloads the calcs entirely so it doesn't have to pay for the cache miss and the misprediction cycles???

Otis, you know way more about the real guts of a CPU ... what say you?

side note: manual intervention to get double PAOLAs is a royal pita

I know what I'm talking about? Oh brother...

I would suspect that most of the improvement in times is coming from the GPU not slowing down...I'd say you're looking at a 13% improvement, is the GPU changing 13% in clock speed?

If I had to pick something else I'd say you're running his on a machine with HT right? My guess would be that when you are only running one task the CPU is going into a low power state as it sees there is very low utilization on a very low priority task. I would suspect that it's turning all but 1 core off and then using that core to do all of the work for this unit as well as the work needed for the OS and other stuff. When you add, I would guess, 6 other CPU tasks and give each of these their "own" cpu, even though it claims it's only using 1 CPU total, I suggest it's playing some games. It's running 6 CPU tasks across the first 6 CPUs, then instead of running these on the same core, it's giving each of them their own "half-core" which is actually significantly better as there's less context switching going on. The work the OS and other things are doing on the computer is then shared between 8 loaded cores instead of 1 so it has less of an impact.

But that's all speculation just on my understanding of what you're saying. There is a very valid argument to say that cache is playing a role, but that would depend upon implementation details that I don't know of the top of my head (if they are even released) and the rule of thumb is more cache never hurts... I've never actually seen a real world case where added cache decreased performance. (sure there are cases involving conflicting collisions evicting data from the memory right before it is used, but that is a very rare occurrence and to my knowledge does not actually occur in practice with modern computers)

But then again, I'm quite rusty on my computer architecture... I'd have to sit down for a while and think it through to give you a better guess...

@werdwerdus