GPUGrid have determined that they are underclaiming credits, and will adjust in future version of the client. Also, they have decided to award a 50% bonus for results returned within 24 hours (see thread (http://www.gpugrid.net/forum_thread.php?id=776&nowrap=true#7832)). This is touch and go on my 9800GT which has been taking over 12 hours for a WU and GPUGRID fetches 2 WUs, but luckily I got a short WU (10 hours) and got a 50% bonus on the last WU returned. They may increase the bonus time limit from 24 to 36 hours which would be great. In the meantime a bit of babysitting.
No doubt this decision will provoke some lively discussion!
wow, that is alot of credit!! NOICE:)
I still dont get exactly how these GPU work units are actually "crunched"..
I heard someone hereon our forum say that the time shown in Boinc-manager is only the amount of time that a CPU spends on the WU....
Why then doesn't suspending all other work except GPU result in the WU just ticking over like a normal WU from anyone else ?
I ask this as Dyeman's post seems to say that all "batches" (ie his 2 that GPU grid fetches at a time) must be returned in that time. I think mine grabs two at a time as well, but it does it even when I have it set to "no new work", and basically does it anytime I "update" GPU grid on the BM.
Any plain English explanation would be appreciated.
Thanx
GPU WUs are crunched by the GPU rather than the CPU, though they do use a small amount of CPU time to "feed" the work to the GPU. It tends to be this CPU time that shows in the BOINC Manager, so it in no way reflects the actual time that the work takes. Because the GPUs have many cores they can crunch a lot more work than the CPU can in any given time period, hence the credit for GPU work appears much higher than for CPU work. The GPU tends to run at capacity, so disabling other work through the CPU will not influence the crunching time for GPU work.
The BOINC Manager seems to be programmed to download GPU work according to how many CPU cores a computer has. This is not particularly logical but is the current system though I expect that this will change as the system evolves and newer BOINC versions are released. What it means currently is if you have the card in a single core or dual core machine, you are more likely to be able to return the work within the "bonus" period than you are on a quad core or i7, unless you micromanage it closely. My computers do NOT download GPU work when I have it set to NNW, and would be surprised if yours does. It should not happen, so if it does it is probably a bug in your BOINC version - I still use 6.4.5 and it works correctly.
Yes, it's prtobably a bug.
However, setting a small cache helps too. I am running GPUGrid on a Q6600 with 9800GT card and usually only have two GPUGrid WU in the system. That's with a cache of about 0.1 days
For me. GPUGRID seems always to try to keep 2 WUs in the cache - when one is uploaded a replacement is downloaded. Thhis is with a 2 day cache (Settings: Computer is connected to the Internet about every 0 days, Maintain enough work for an additional 2 days). So if the card takes 12 hours and 1 minute to process a WU, the elapsed time for each WU would be 24 hours and 2 minutes and miss out on the bonus. At 11 hours and 59 minutes it would be OK. If they increase the limit for the bonus to 1.5 days the there should be no problem.
In the interim, setting NNT until a Wu is well progressed before manually downloading the second WU should ensure getting the bonus credit (I hate babysitting.. :-\)
Done lots of overtime this week so I might look at getting a GTX260 soon ;D
A GTX260 will have no problem meeting the 24 hour deadline !! ;D
what about the 250 ?? that's about my spend limit ATM...
I'll let ya know when the AA finishes. Will try the GTS 250 in TheGnat first, and then iGnatious
GTS 250 is essentially the same as a 9800 GTX+. It has a few more shaders than 9800 GT and they are significantly faster, so I reckon it should have no problem meeting the 12 hour elapsed time required to beat the 24 hour deadline (and certainly no issue if the deadline is increased to 36 hours)
Don't ya just love it when the latest file downloaded from GPRGrid says that it has a completion time of 48 hrs?? Not even a 280 could process that within 24 hrs *GRRRRRRRRRRRRRRRRRRRRRRRRRRRRR*
Quote
To easy the effort, we are giving now:
60%+ for WUs returned within 1.5 days
20%+ for Wus returned within 2.0 days
We will probably also extend the deadline to 5 days for all the others.
gdf
Update on the deadline thresholds.
Quote from: Furlozza on March 28, 2009, 10:01:51 AM
Don't ya just love it when the latest file downloaded from GPRGrid says that it has a completion time of 48 hrs?? Not even a 280 could process that within 24 hrs *GRRRRRRRRRRRRRRRRRRRRRRRRRRRRR*
Hey, I have been getting some with an estimated completion time of 110 hours! Go figure how you return that within the deadline of 4 days! Mind, you I did manage to process one of these in 110 hours, and even though it was half a day beyond deadline it still got credit. biggrin
WS
I trust the reward was worth the effort? That is, more than 4000 credits per WU
4214 on a 9500GT. biggrin
http://www.gpugrid.net/result.php?resultid=428797
Mine ALWAYS shows an estimated time of more than a thousands hours.
But the longest one took a little over a day
Running BOINC 6.5.0 (64bit Linux) on that box. Nothing in the cc_config file as far as # of processors goes but it's always running 1 GPU and 4 others. GT9800 card
Quote from: Wang Solutions on March 28, 2009, 12:22:35 PM
Quote from: Furlozza on March 28, 2009, 10:01:51 AM
Don't ya just love it when the latest file downloaded from GPRGrid says that it has a completion time of 48 hrs?? Not even a 280 could process that within 24 hrs *GRRRRRRRRRRRRRRRRRRRRRRRRRRRRR*
Hey, I have been getting some with an estimated completion time of 110 hours! Go figure how you return that within the deadline of 4 days! Mind, you I did manage to process one of these in 110 hours, and even though it was half a day beyond deadline it still got credit.
110 hours :thumbdown: I thought it was bad enough getting a 26 hour WU. Must be part of a cunning plan not to pay the bonus biggrin
update - looks like it will run in about 14 hours - BOINC estimate a bit wonky again
Quote from: Wang Solutions on March 28, 2009, 12:42:00 PM
4214 on a 9500GT. biggrin
http://www.gpugrid.net/result.php?resultid=428797
I have a 9500GT in one of my linux boxes, but cannot get GPU grid running...
heeeelllpp !!!!!
Quote from: veebee on March 28, 2009, 10:55:51 PM
Quote from: Wang Solutions on March 28, 2009, 12:42:00 PM
4214 on a 9500GT. biggrin
http://www.gpugrid.net/result.php?resultid=428797
I have a 9500GT in one of my linux boxes, but cannot get GPU grid running...
heeeelllpp !!!!!
It will be slow, about 3 days on a long wu.
Use 177.82 drivers with 6.4.5 boinc client.
Are you on Ubuntu?
Yes, Ubuntu 8.10.
I think its an issue with "libcudart".. though I just put a copy of it in /var/lib/boinc-client/ which is where I think its supposed to be ... still no good after a restar of client..
I kno wit'll be slow, bu tit's better than not crunching at all with the thing... - though, I suppose I could do F.A.H.
Methinks me needs to get a GT250 minimum.
Quote from: veebee on March 28, 2009, 11:24:54 PM
Yes, Ubuntu 8.10.
I think its an issue with "libcudart".. though I just put a copy of it in /var/lib/boinc-client/ which is where I think its supposed to be ... still no good after a restar of client..
I kno wit'll be slow, bu tit's better than not crunching at all with the thing... - though, I suppose I could do F.A.H.
Methinks me needs to get a GT250 minimum.
8800GT is good for two wu's a day.....good card for the money if you can find them used.
I run 8.04. Just put your boinc client in a folder in your home folder. Run it from there. It will have to be started manually[boinc manager] each time unless you go to "sessions" and have it start there on boot[boinc clent and boinc mgr].
has started running overnight... seems to be the story - people (not all obviously) hook up to GPU grid and their comp "doesn't recognise cuda devices"... then a while later.. VOILA ! its crunching GPU grid !
MAGIC !!!!
Be careful with 8800GT - I think that the earlier ones using the G80 chip don't support the level of CUDA (1.1) used by GPUGrid - see (see this thread) (http://www.gpugrid.net/forum_thread.php?id=873).
Also regarding GTS250, I think GTX 260 may be better bang for the buck if you can stretch to the extra $$. Have seen them for just over $300 and should crunch GPUGRID significantly faster than a 250/9800GTX+. A 260 is nearly twice as fast as a stock 9800GT (in the area of 7 to 8 hours per WU).
Also in terms of future apps, keep in mind that the 250 is really previous generation technology (essentially a faster clocked dye shrink of an 8800 GTS) and doesn't do double precision floating point, unlike the GTX 260 and above. Milkyway for example requires double precision FP, which is the main reason the high-end ATI cards (which have very strong double precision FP) have a client, and CUDA does not. When a MW CUDA client does happen I expect it will only support GTX 260 and above, not the 250 or any other 8 series or 9 series cards.
Quote from: dyeman on March 29, 2009, 08:10:25 AM
bbbbbbbbbbBe careful with 8800GT - I think that the earlier ones using the G80 chip don't support the level of CUDA (1.1) used by GPUGrid - see (see this thread) (http://www.gpugrid.net/forum_thread.php?id=873).
Also regarding GTS250, I think GTX 260 may be better bang for the buck if you can stretch to the extra $$. Have seen them for just over $300 and should crunch GPUGRID significantly faster than a 250/9800GTX+. A 260 is nearly twice as fast as a stock 9800GT (in the area of 7 to 8 hours per WU).
Also in terms of future apps, keep in mind that the 250 is really previous generation technology (essentially a faster clocked dye shrink of an 8800 GTS) and doesn't do double precision floating point, unlike the GTX 260 and above. Milkyway for example requires double precision FP, which is the main reason the high-end ATI cards (which have very strong double precision FP) have a client, and CUDA does not. When a MW CUDA client does happen I expect it will only support GTX 260 and above, not the 250 or any other 8 series or 9 series cards.
New list of cards here;
http://www.gpugrid.net/forum_thread.php?id=867
The 8800"GT"/"GS" is ok,8800 GTS/GTX/ULTRA are the ones to avoid.
Does that suggest the GeForce GTX 295 at $AU771.00 is the best one to use?
Quote from: Mike Mitchell on March 29, 2009, 09:05:06 PM
Does that suggest the GeForce GTX 295 at $AU771.00 is the best one to use?
I wouldn't say so unless you have a PSU that can handle the load... And it depends on whether you want one or more GPUs.
That's one of the reasons I'm considering getting a 9800GT (around $170-180 each) in each of my quads, rather than a single GTX 260 (around $330) in one of them.
Put another way. If I were to put a GTX295 in one of my boxes, I'd have to also spend upward of $200 to get a decent PSU (like a Corsair HX620, Zalman ZM600, Coolermaster M620 etc)
While I have two decent power supplies I didn't really want spend $AU771.00 but it's nice to dream. biggrin
Yeah, nothing wrong with dreaming. Get a mobo with 4 PCI Express x16 slots and put a GTX295 in each :) Would have to have an i7 CPU at this point (or a dual socket quad core Xeon/Opteron) to get work for each GPU!
Now that's the way to go. If I win tattslotto I'll be straight into it.
Hmm bonus credit for being fast now that would go down great for my 280 and 295 lol. I got the cache sorted on my systems finally and i only very occasionally have a spare GPU WU. Thats generally when one is almost done. Im real tempted to add another GTX 295 now.
Edit: Scratch all that the servers playing up again and not crunching GPU stuff......grrrrrrrrrrrrr