CUDA 2.1 discussion

CUDA 2.1 is now available–please post feedback in this thread.

The link of

CUDA SDK

CUDA Visual Profiler

CUDA Debugger

is broken now.

yeah, I just noticed that–trying to figure out what’s up…

Also the way the URLs are displayed under toolkit and SDK section makes it difficult to go get the thing for your platform…

Like, you cant tell just by seeing what is xp, what is xp 64-bit, what is fedora rhel 4 or whatever…

Does CUDA 2.1 has support for:

  1. MCUDA stuff - Compiling for multiple CPUs?
  2. Querying if a CUDA card has 5-sec Watchdog timer or not?

Bug with the opengl driver on windows xp 32/64.

Windows API “DescribePixelFormat” can not work.

Please try the microsoft OpenGL sample.

http://msdn.microsoft.com/en-us/library/ms970779.aspx

I have a standard 64-bit 10.3 Suse linux system and am using a gtx 260. I tried installing the 180.22 driver, but alas, the nvidia kernel model doesn’t want to load.

dmesg shows:

NVRM: loading NVIDIA UNIX x86_64 Kernel Module  180.22  Tue Jan  6 09:15:58 PST 2009

NVRM: The NVIDIA probe routine was not called for 1 device(s).

NVRM: This can occur when a driver such as rivafb, nvidiafb or

NVRM: rivatv was loaded and obtained ownership of the NVIDIA

NVRM: device(s).

NVRM: Try unloading the rivafb, nvidiafb or rivatv kernel module

NVRM: (and/or reconfigure your kernel without rivafb/nvidiafb

NVRM: support), then try loading the NVIDIA kernel module again.

NVRM: No NVIDIA graphics adapter probed!

NVRM: The NVIDIA probe routine was not called for 1 device(s).

NVRM: This can occur when a driver such as rivafb, nvidiafb or

NVRM: rivatv was loaded and obtained ownership of the NVIDIA

NVRM: device(s).

NVRM: Try unloading the rivafb, nvidiafb or rivatv kernel module

NVRM: (and/or reconfigure your kernel without rivafb/nvidiafb

NVRM: support), then try loading the NVIDIA kernel module again.

NVRM: No NVIDIA graphics adapter probed!

PCI: Setting latency timer of device 0000:01:00.0 to 64

NVRM: loading NVIDIA UNIX x86_64 Kernel Module  180.22  Tue Jan  6 09:15:58 PST 2009

NVRM: The NVIDIA probe routine was not called for 1 device(s).

NVRM: This can occur when a driver such as rivafb, nvidiafb or

NVRM: rivatv was loaded and obtained ownership of the NVIDIA

NVRM: device(s).

NVRM: Try unloading the rivafb, nvidiafb or rivatv kernel module

NVRM: (and/or reconfigure your kernel without rivafb/nvidiafb

NVRM: support), then try loading the NVIDIA kernel module again.

NVRM: No NVIDIA graphics adapter probed!

I gave the system a reboot just to be sure, but the new driver is a no-go for me so far. Manually loading the nvidia kernel module failed as well. I don’t believe the modules complained about above are plugged in on my system; lsmod doesn’t show 'em. Dutifully reporting, B.C.

@BoxedCylon,

What I can infer from dmesg is that: Some other driver has already taken posession of the device. Is the previous driver still running?

  1. no

  2. absolutely does.

BoxedCylon, do you have drivers from the SUSE repo installed?

Any news on the links being broken? I have a small window of opportunity to install an update now, otherwise it will become next week or later :">

Not yet, I expect I’ll hear something in the next 10 hours or so. Then again, install the driver and the toolkit (as far as I know those aren’t broken), install the SDK to a home directory later?

Yeah, I’m stupid. Those downloaded well. That will teach me to not open 20 tabs in firefox ;)

I don’t think so - X11 wasn’t running and I checked that the previous nvidia module was removed. And then I also rebooted just to be on the safe side, and X11 still did not start. I am, of course, an old hand at all this stuff (which also means I can appreciate that sometimes I can make remarkably stupid mistakes sometimes…).

Hooray! Figure out where it’s broken (I haven’t seen anything yet besides the Volkov FFT bug which we’ve posted the workaround for), let me know :)

I have a stupid question. I always check the NVIDIA X server settings tool to check if I have loaded the appropriate driver. I notice that the slowdown threshold for FX4800 is 190 degrees celsius. Is that correct? (I never checked it before, but is seems quite hot ;)).

I’ll start to break the release now, although I am not often able to break it…

get~~~

That might be a bug; I’ll check that tomorrow.

Running a brand-new kernel that probably has an infinite loop on my display device (FX4800) from within MATLAB does not trigger the watchdog somehow. I had to hard-reboot the system (waited for 3 minutes).

One additional bit of information. When in console mode and I attempt to load the nvidia module by hand I get:

FATAL: Error inserting nvidia (/lib/modules/2.6.22.19-0.1-default/weak-updates/nvidia.ko): No such device

Too bad.

lspci reports:

01:00.0 VGA compatible controller: nVidia Corporation Unknown device 05e2 (rev a1)

for my rapidly devaluing gtx 260.

Everything is running OK here on my gentoo box. I’ll upgrade the servers tonight and post if there are any troubles.

Where did the driver performance improvements from CUDA 2.1 beta go? HOOMD was 10% faster on that (compared to CUDA 2.0) due to optimizations that reduced the driver overhead for very short kernel launches. HOMD on CUDA 2.1 is now back to the same performance as it was on CUDA 2.0.

Also, are the linux64-rel-nightly-2.1-gobblygook#s filenames the permanent file names for the release? Or are they going to be simplified into the same naming scheme as used for the 2.0 release?