I might try that over the weekend.
I did set my GTX260 clocks back to stock and tried a 6.72 task on that card but it failed immediately. I have another couple of GT240s that I can set to stock and try first - I dont want to mess too much with the quad system as there are always tasks running at different stages.

- Just checked and the +12V rating is 41A (492W) for the Corsair 550.
There is no way I am pushing that with 4 GT240s and a stock CPU.
The CPU has a 125W TDP and the GT240s are 69W TDP.
The one HDD, one DVDRW, and 2 sticks RAM wont do much damage either.

- Just measured it on a Watt metre and the whole system is only drawing 235W when crunching (CPU 90%, 4 GPUs at about 75%).

No problem with WUs at different stages, it'll just suspend one until a GPU becomes available. As far as the power, a big question is whether the MB can keep the power supplied properly to all 4 PCIe slots with no additional power supplied by PCIe connectors. It's not like you're running HD4770 cards that draw not much more than the GT240 but supply most of it through a PCIe connector to each card.

Suspended all tasks, set clocks to default (system power usage fell to 140W), enabled the one 6.72 task to run (no other CPU or GPU tasks were running).

The 6.72 task crashed in 5sec!

I think the problem is not power related:
When these GPUs are not crunching they only use about 10W (actually about 6W, with the board drawing the other 4W).
So, with the other cards not crunching (and the CPU not crunching) the cards only used about 90W between them.
90W can’t be an issue, because if it were the cards could not crunch four 6.03 tasks at the same time; using about 240W between them!

It might be the fact that I am using more than one card, some Boinc setting, the WU’s, or Vista?

The WUs that you are reporting that error out quickly all throw the same error:

ERROR: file ntnbrlist.cpp line 63: Insufficent memory available for pairlists. Set pairlistdist to match the cutoff.
called boinc_finish

this is what the WU is reporting you have for memory on each of these cards.
# Total amount of global memory: 497614848 bytes

I took a look at the machines that passed these same WUs and did find a 240 that reported a different amount of memory and it passed:
# Total amount of global memory: 536870912 bytes

so maybe the WU is early terminating not because it threw an error as much as it was smart enough not to even try calculating the real science.

I know that memory requirements for the WUs have been an issue in the past and while we thought we were past that I think it needs to be revisited.


Does anyone know if there is one version of BOINC that is more accurate than others at reporting available GPU memory?

Do the apps rely on BOINC to say how much free memory a GPU has?

Can the app get a more accurate number than BOINC?


In thinking about not only the insufficient memory issue but also guessing that there could be substantial performance increases if you used more of the available memory on any given card (e.g. a 480 has 1.5 gb but the max usage I have observed is 437 mb) I can think of two approaches that may be worth consideration.

Could the feeder be made more "memory aware" of the machines as they ask for work to parcel out WU more appropriately? I know there will be machines that have multiple GPUs that are not exactly matched but even if you target the smallest amount of memory card for the machine asking for work, at least it would not fail due to lack of memory. I am also guessing that the same people who run multiple card systems are the same people who will actually understand this and optimize their systems accordingly.

My other thought, which may be even more difficult to implement, is to make the apps themselves more dynamic in their ability to determine and allocate available memory.

---

Thanks - Steve

BOINC developers have been struggling with CUDA memory questions recently. They were trying to grapple with two different measures:

1) How much actual memory does the card have?
2) How much space does the science application have available to run in, after system overheads (notoriously, Aero, and likewise OS X) have taken their bite.

The live, running memory detection in particular was causing problems, so they've taken that bit back out again in the latest v6.10.55/.56: but you could probably try with .54 or just before if you want to see what happens.

The other noticeable thing is that the 197.xx drivers tend to give a smaller answer for total memory size than earlier drivers. Did you compare driver versions on those two 240s with different global memory reports?

The cards I have are all 512MB GT240 DDR3, according to the manufacturer and GPUZ, so I am not sure what or why they are being reported as having 474.5625MB

With the recommended NVidia driver (197.45) and Boinc Version (6.10.18):
15/05/2010 13:09:08 NVIDIA GPU 0: GeForce GT 240 (driver version 19745, CUDA version 3000, compute capability 1.2, 475MB, 307 GFLOPS peak)
Boinc 6.10.45 & NVidia 197.45:
15/05/2010 13:11:26 NVIDIA GPU 0: GeForce GT 240 (driver version 19745, CUDA version 3000, compute capability 1.2, 475MB, 307 GFLOPS peak)
Boinc 6.10.56 & NVidia 197.45:
15/05/2010 13:11:26 NVIDIA GPU 0: GeForce GT 240 (driver version 19745, CUDA version 3000, compute capability 1.2, 475MB, 307 GFLOPS peak)
Boinc 6.10.56 & NVidia 197.57 (beta):
15/05/2010 13:47:28 NVIDIA GPU 0: GeForce GT 240 (driver version 19757, CUDA version 3000, compute capability 1.2, 475MB, 312 GFLOPS peak)

The card you found reports exactly 512MB. What Version of Boinc & NVidia driver was it using and what operating system (as there are different drivers for different operating systems)?

I also noticed that my GTX260sp216 (in a different system) has less than 1GB RAM; 881MB according to Boinc, and 896 according to GPUZ.
This card also fails 6.72 tasks when at stock, but crunches the 6.03 tasks (it’s a good card), using recent drivers too.

I will try the NVidia 196.21 drivers.

This is the WU I saw where someone else's 240 completed but your's said insufficient memory:
http://www.gpugrid.net/workunit.php?wuid=1465191

They do a have different driver: 19562 on Microsoft Windows 7 Ultimate x64 Edition, (06.01.7600.00).

Your 260 is getting a different error:
http://www.gpugrid.net/result.php?resultid=2330409
- exit code -40 (0xffffffd8)
SWAN: FATAL : swanMalloc failed

I think it is also a memory issue, I just do not know enough to guess as to what, specifically, it did not like .

Hopefully looking at these instances will help GDF and crew come up with a solution.

---

Thanks - Steve

15/05/2010 16:05:31 NVIDIA GPU 0: GeForce GT 240 (driver version 19621, CUDA version 3000, compute capability 1.2, 512MB, 257 GFLOPS peak)
15/05/2010 16:05:31 NVIDIA GPU 1: GeForce GT 240 (driver version 19621, CUDA version 3000, compute capability 1.2, 512MB, 257 GFLOPS peak)
15/05/2010 16:05:31 NVIDIA GPU 2: GeForce GT 240 (driver version 19621, CUDA version 3000, compute capability 1.2, 512MB, 257 GFLOPS peak)
15/05/2010 16:05:31 NVIDIA GPU 3: GeForce GT 240 (driver version 19621, CUDA version 3000, compute capability 1.2, 512MB, 257 GFLOPS peak)

Just need to wait for a 6.72 task to come my way, there are none at the minute.
I also put 196.21 onto the system with the GTX260.

I guess Boinc ver 6.10.45 (and variants) might have worked well right up to the point where people started using the 197.xx drivers!

Will there be any 6.72 tasks available over the weekend?

Well, I picked up two 6.72 WUs on the GT240 system.
So far (1h 15 min), so good – Using Vista 64bit, 196.21 drivers and Boinc 6.10.56
- Note they use to fail after about 7sec (197.45 drivers)

It might be a Driver/System/Boinc combination problem. I suppose if it was easy to spot it would have been fixed by now.

6.72 tasks seem to work fine for the 19745 driver for XP x86 on Boinc 6.10.51:
2309288 1457946 10 May 2010 17:37:28 UTC 11 May 2010 11:40:02 UTC Completed and validated 60,352.50 5,833.84 7,954.42 11,931.63 Full-atom molecular dynamics v6.72 (cuda)

But there are differences between XP and Vista drivers.

It looks like the problems (here) are just with Vista & W7 drivers when it comes to 6.72 WU’s.

Anyone else with 6.72 issues who uses Vista or Win7 – Use a 196.xx driver (196.21 seems to work).

Snow Crash, Beyond and Richard Haselgrove,
Thanks for your help with this.

So far, three 6.72 tasks completed on the Vista system with the GT240s.
There have been no failures (since changing the driver to 196.21), and 4 tasks are running well (2 should finish within an hour):
2337257 1477078 15 May 2010 21:08:32 UTC 16 May 2010 11:43:47 UTC Completed and validated 32,909.90 3,104.78 4,503.74 6,755.61 Full-atom molecular dynamics v6.72 (cuda)
2336451 1476466 15 May 2010 18:01:24 UTC 16 May 2010 8:33:23 UTC Completed and validated 33,353.63 3,316.39 4,503.74 6,755.61 Full-atom molecular dynamics v6.72 (cuda)
2336422 1476441 15 May 2010 17:56:50 UTC 16 May 2010 3:12:30 UTC Completed and validated 32,999.91 3,178.07 4,503.74 6,755.61 Full-atom molecular dynamics v6.72 (cuda)

So, 196.21 seems to work well for Vista x64 when crunching 6.72 WU's (using Boinc 6.10.56)

Driver 197.45 still seems to be working well on Win XP x86 with Boinc 6.10.51 for 6.72 tasks.

Driver 196.21 might still have issues on Win7 x64:
With Boinc 6.10.50 (GT240):
This one failed with less than an hour to go, but looks like it is a different issue (its not like it failed after 7 seconds),
2335968 1476139 15 May 2010 15:50:19 UTC 16 May 2010 6:40:12 UTC Error while computing 33,138.30 5,189.69 4,503.74 --- Full-atom molecular dynamics v6.72 (cuda)]
This one finished (same system),
2301423 1452813 10 May 2010 17:58:29 UTC 11 May 2010 4:08:09 UTC Completed and validated 35,936.96 5,254.89 4,503.74 6,755.61 Full-atom molecular dynamics v6.72 (cuda)
- changed to native GPU & RAM clocks, with only 1575 to 1600MHz shaders, & upped the fan speeds to see if that makes them more stable with these 6.72 WU’s

With Boinc 6.10.51 for a GTX260sp216 (native clocks):
All still failing after about 8sec, so same problem as before (well actually it was 5sec before):
2337284 , 2337028 , 2332609
- Have moved up to Boinc 6.10.56. If that does not work I will try a different driver.

Thanks,

Can anyone give me a hand with how to set my preferences? I've read through the thread and I am trying to figure out if they can be set so that I only receive the 6.72 WUs on my system. These are really running great on my GT240s so if it is possible to only receive them that would be great.

-Matt

Can anyone give me a hand with how to set my preferences? I've read through the thread and I am trying to figure out if they can be set so that I only receive the 6.72 WUs on my system. These are really running great on my GT240s so if it is possible to only receive them that would be great.

-Matt

6.72 will soon become the production application I suspect. for now you can go to your account page and select GPUGRID preferences. click on "Edit GPUGRID preferences" under "Run only the selected applications" check "ACEMD", uncheck "ACEMD ver 2", check "If no work for selected applications is available, accept work from other applications?" or you may not get work and then click on "Update Preferences" and you're done.

---

Radio Caroline, the world's most famous offshore pirate radio station.
Great music since April 1964. Support Radio Caroline Team -
Radio Caroline

Sure enough, GDF just announced the merge after I posted.
Thanks, Matt