Nvidia GPU problem
23 May 2010 15:17:18 UTC
Topic 194939
(moderation:
I recently signed up for Einstein and when I connected it downloaded a dozen+ WU. I noticed the GPU (ABP2cuda23) were not getting processed, but the standard CPU WU were. I thought maybe the nvidia driver was old, so I got the latest one. (v195.36.24)
Yet still can not get the gpu WU to process, even if I pause all other projects leaving only Einstein as active. I have obviously restarted the machine and BOINC. Here is the line from boinc:
NVIDIA GPU 0: GeForce 9400 GT (driver version unknown, CUDA version 3000, compute capability 1.1, 511MB, 45 GFLOPS peak)
Any ideas other than disabling GPU processing for Einstein?
Thanks
-J
Nvidia GPU problem
)
The only thing you could try is making more memory available for CUDA processing, by lowering display resolution and/or stopping graphic-memory-intensive applications.
For more information on this BOINC bug see this thread (and the ones linked in there).
Gruß,
Gundolf
Computer sind nicht alles im Leben. (Kleiner Scherz)
RE: I NVIDIA GPU 0: GeForce
)
Gundolf's advise is valid, but I really doubt whether you really want to run E@H CUDA jobs on your 9400 GT. As indicated by the benchmark (45 GFlops peak), this card is at least 10 times (!) slower than mid to high-end cards. It has only 16 processing cores, while it's more standard now to have 256 or more cores, and at a higher clock rate.
So the cruel fact is that ABP2 CUDA jobs would almost certainly run slower on your GPU (plus one CPU core is needed as well!) than on your CPU alone.
I have a notebook now with a not_so-fast (but power-saving) mobile graphics card, and it has only 48 cores, three times more than your card. It crunches ABP2 jobs just a bit faster than on the CPU alone (in this case a hyperthreaded i5 core).
So I guess the break-even point on most hosts where you want to start considering ABP2 CUDA jobs is somewhere between 32 ... 48 GPU cores.
Sorry for the bad news
HB