CUDA 2.0beta2: new release... old bugs

Hello guys,

I've been eagerly waiting for a new release of CUDA, with the hope that some annoying bugs had been resolved. Well, a new release is out, but the bugs are still there.

I’m starting to be quite frustrated by this fact. I’m using CUDA for my degree thesis, and since a few weeks I’ve essentially stopped development because of these bugs. CUDA technology looks very promising, NVidia is investing much effort on it, and so do I, but still these bugs sometimes just vanify my will to use it.

I’ve just installed che driver and the toolkit of CUDA 2.0beta2 on a fresh 7.10 Ubuntu install. First of all, there is the driver installer bug that prevents the installation to be fully functional: if I install the driver and reboot, the X server falls in low-graphics mode and I have to reinstall the driver to make it function properly.

Then, the nasty bug I’ve been reporting is still there, worse than before.

I’m not attacking NVidia and the like. But how can I effectively work with CUDA, when I can’t even be sure that my compiler is correct?

Yours sincerely,

This sounds like a problem I have encountered too. In my case, it was Ubuntu’s software / driver management system that interfered with the usual installation of NVIDIA drivers.

Hope you have good luck with the other stuff.

I’m sorry to hear that you’re experiencing difficulties. With respect to the X driver, did you report that problem per the instructions in the X driver README?

As for the CUDA app problem, please generate and attach an nvidia-bug-report.log from your system, along with the output from “nvcc -V”.

I should note that there are currently known issues when using CUDA with GPUs that have less than 256MB of framebuffer. Does the SBIOS in your notebook allow you to configure the framebuffer allocation to a value greater than 128MB?


I’m very glad to hear feedback from you. Excuse me for the harshness of my first post, I was a bit frustrated :-) I’m investing quite much in CUDA: I’m using it for my thesis on Super Resolution algorithms, for my ray-tracing experiments, and I’m also writing a .Net object-oriented wrapper for it (of course I’m releasing it as soon as I can).

As far as I know, my notebook has 128 MB of DEDICATED video memory. The graphics device uses the main RAM if the dedicated one is not enough: accordingly, my NVidia Panel says: “Dedicated video memory: 128 MB; Total available video memory: 895 MB”.

Well, I’m investingating into this. Thank you again for the feedback.

Ubuntu’s built-in NVIDIA drivers are interfering with your new one; this is because Ubuntu’s package system doesn’t know that you’re trying to use another driver, so it didn’t disable the old one. To fix, add the following line to /etc/default/linux-restricted-modules-common:

DISABLED_MODULES=“nvidia nvidia_legacy nvidia_new”

This prevents the loading of the old drivers, allowing your new driver to be used without interference.

Well, it seems that CUDA sees only the dedicated memory. This screws me up since I’ve got 128 MB of such a memory (even if Windows show that I’ve got 895 “total” memory).

Argh, I hope this issue is resolved since I can’t afford to buy a new computer…

Quick suggestion -

Try to get a cheap development box w/ a cheap 8000 series card. An old P4 will do. You will be able to do this for very cheap. Hook it up to the network, use VNC or NX to connect, and develop away.

I will build you a statue for this.