Latest drivers (NVidia and ATI)

[Richard Haselgrove]

I never explored the possibilities with the 470. Was running stable with 266.58, and I left it at that. (I don't really like running Beta drivers, along with Beta apps - change one thing at a time, I reckon).

Having said that, it threw a complete wobbly and crashed BOINC, trashed SETI tasks, about an hour ago. Maybe heat - first really warm day of summer. I'm giving the machine a checkover, loading 275.33, will probably start afresh with x38b when I'm sure it's working OK. Some interesting error messages to post when I get control of the machine again.

[Claggy]

My host threw a wobbly yesterday, kept locking up, it then managed to loose the Einstein project, after reatttaching to Einstein and getting the same Wu's back, it did it again,
so i disassembled it, disassembled the GPU's and cleaned them out, reassembled them, then cleaned out the CPU cooler, now stable again, phew,

Claggy

[benool]

I tried 275.33 today and had to revert immediately as tasks were all completing with 'computation error' within a few seconds.

Looking at BOINC initialization messages, I think it could be related to the new unified memory addressing introduced in CUDA 4.

With 266.58 I get this:
03/06/2011 11:38:31      NVIDIA GPU 0: GeForce 8600 GTS (driver version unknown, CUDA version 3020, compute capability 1.1, 240MB, 96 GFLOPS peak)

With 275.33 I got something like this (not exact copy/paste):
NVIDIA GPU 0: GeForce 8600 GTS (driver version unknown, CUDA version 4000, compute capability 1.1, 209MB, 96 GFLOPS peak)

Notice how suddenly there seems to be only 209Mb available out of 256Mb from the card?

If memory serves well, MB cuda takes around 220Mb so 240Mb is already just barely enough but at 209Mb it just isn't enough anymore and it fails...

[Miep]

x38e JUST runs on my 256 MiB card on 266 drivers with 253 MiB used in total. 270 driver goes over the edge, memory need wise (same for x32f). I haven't tried 275, but I don't think the memory footprint has gotten smaller again. Until such time as Jason can reduce required memory, I fear 256 MiB card owners are stuck on 266.58 drivers.

Thanks for confirming that the are other 256 MiB cards out there beside my Quadro 570M alpha testing system.

[benool]

You're welcome.

I forgot to mention it but yes, the tests were made with x32f version.

[perryjay]

My little GTS 450 shows this with the 266.58 driver...

6/3/2011 8:23:51 AM      NVIDIA GPU 0: GeForce GTS 450 (driver version 26658, CUDA version 3020, compute capability 2.1, 993MB, 461 GFLOPS peak)

I tried the newer drivers before and though it didn't break any work units it did slow way down and I had to go back to this 266.58. I was so worried by the slow down I didn't think to see how much the memory changed. I sure will be glad when this problem gets solved.

Oh yeah, I'm also running the X32F.

[Miep]

[grumpy] Does anybody EVER read anywhere properly?[/grumpy]

Until either x37/x38 is available puiblicly OR the boinc API has been updated AND stock recompiled,  chances are high that Cuda 4 drivers (270+) wil NOT work properly. sticky downclock, memory issues, strange exits truncaed stderr, you name it. Stay on the stuipd 266 driver until we have the installer out.

You can of course just riosk it and hope your system isn't affected but you have been warned. Several times.

Entschuldigt, aber ich bin es einfach leid mir den Mund fusselig zu reden, nur weil keiner zuhoert. Irgendwann reisst auch mir der Geduldsfaden.

I think that's too colloquial to translate well with google. sorry.

[perryjay]

Key phrase there was I tried the newer drivers before.... That was when they were first released. I was saying I was not going to update drivers again until this mess is sorted out. That's my round about way of agreeing with you "Grumpy" !!   

[Pepi]

When you have card like me with 1 GB of memory , then you are not worried about that info ( what drivers say about memory size left) I know that on nearly any new release of drivers show less memory: I have GTX 460 with 768 MB of memory, and in that case you cannot run 3 only 2 WU in parallel. But when you have  1 GB, it is not problem what drivers show ( 903 MB, 935 MB or 993 MB)
But this drivers is fast as hell!

[Miep]

Key phrase there was I tried the newer drivers before.... That was when they were first released. I was saying I was not going to update drivers again until this mess is sorted out. That's my round about way of agreeing with you "Grumpy" !!   

I do apologize. wasn't my day today.

[perryjay]

No problem Carola.

 And Pepi, My GTS 450 also has 1Gb but it didn't take long for the newer drivers to cause it to down clock. Probably because I snoozed the GPU to play a game and it wouldn't come back up. Some people were reporting it took over 30 days before theirs down clocked but down clock they did. I sure am looking forward to the new installer and a cure for it. I will be glad to jump on it when it comes out as Beta. Wish I knew enough to join the alpha testers but I would just slow them down!!   

[Pepi]

Perryjay, snozing GPU is main reason for downclock. I make another tactics: I use onboard GPU for 2D work in windows, so Geforce is only for  crunching. 

[perryjay]

Great idea Pepi but my cheapo machine disabled onboard GPU when I put the extra in there. Can't find any way to get it running again with the 450 in there.

[Pepi]

Great idea Pepi but my cheapo machine disabled onboard GPU when I put the extra in there. Can't find any way to get it running again with the 450 in there.

And your motherboard is?

[perryjay]

EMachines  MCP73VT-PM 1.0  Like I said, a cheapo!