well, it seems that some of our previous problems stemmed from our not RTFM; we had mismatched versions on three things in the nvidia stack, all of which need to always be updated together:
in /home/student/Downloads/ on both cudak4 and cudak5 we now have the newest driver (newer than was available from nvidia download page a few days ago when I reported last)
NVIDIA-Linux-x86_64-256.35.run [version 256.35]
we also have a toolkit cudatoolkit_3.1_linux_64_fedora12.run, and finally
the SDK in gpucomputingsdk_3.1_linux.run
all these things except the last need to be installed by a superuser like
#sh (file name with .run extension)
I think you should set up a directory, either NVIDIA_GPU_Computing_SDK/
or like I do, cuda31 (to have cuda toolkit/SDK version visible) **in your own**
home directory. Then you can modify the sdk programs in the, say,
/home/your-name/cuda31/C/src/ directory at will, and use their makefiles which
put the demos or your own program executables in the /home/your-name/cuda31/C/bin/linux/release/ directory
* * *
It seems that all the sdk demos are working properly on all cards, on both
cudak4 and cudak5 now!
[I can see temperatures in the nvidia-settings utility on k4 now, but not yet on k5,
I'll try rebooting. in any case, I have stress-tested k4, and it works ok, also doesn't get stuck in high-temp state].
I'm curious about the CPU speed. the Fedora 12 utility shows that it's remaining in the slow/cool state 1.6(?)GHz, even when running multiple cuda examples. I'd think it should switch to the max ~2.86 GHz. does that have to do with any benchmark results?