Lack of support for 9800GTX in 2.0 Beta

I notice the Beta 2.0 post in the announcements forum mentions that although the 9800GX2 made it in, 9800GTX did not. I’ve got a couple of questions about this:

  1. The lack of 9800GTX support, is that an absolute “this will not work, we are 100% sure” or is it a, “it may work, but try at your own risk, we aren’t supporting you” situation?

  2. Assuming the former of the two situations, is there any timeframe on either an updated beta or final release (whichever is sooner) ?

Many thanks,

Adam

Currently, there are no plans to release any additional beta packages before the final release. All cards available for sale at the time of the final release will be fully supported (including the 9800GTX).

That semi-answered 2), but as for 1), I just decided to work it out for myself. After a little bit (2 lines) of hackery with the nvdisp file, the driver installs for the 9800GTX on Vista 64, and CUDA works fine so far.

Adam

Awesome! Can you run the bandwidthTest example in the SDK and post the results for the Device to Device copy? (I’ve been trying to find someone with a 9800 GTX and CUDA to do this test.)

Quick Mode

Device to Device Bandwidth

.

Transfer Size (Bytes) Bandwidth(MB/s)

33554432 46465.9

EDIT: I expect the figures for the 9800GTX are going to be slightly disappointing when compared to the 8800GTX as it uses a 384 bit memory bus rather than 512? I’ve seen figures on here of around 70GB/s for 8800GTX’s I believe.

Wow, that’s lower than I had expected. The 8800 GTX has a theoretical bandwidth of 86.4 GB/sec, and gets 70 GB/sec device-to-device bandwidth in practice (81%). The 9800 GTX has 70 GB/sec theoretically, but only gets to 66% of that in practice. (?!)

Thanks for running the test. This is very helpful for us.

Although it sounds unlikely, there is obviously a chance that this driver really wasn’t written/optimised in such a way as to support the 9800GTX and it may improve, but you’re right, that is a fair bit lower than the quoted 70.4GB figure quoted on the Nvidia website.

OK, I’m going to have to strongly recommend no one else try what I did and coaxing this driver into installing for the 9800GTX however tempted you are. While it did work, I went to play Trackmania Nations Forever just now and the driver repeatedly crashed and recovered a few times until it eventually BSOD’ed. On the next boot, although Windows booted fine, I got no display output on either monitor, yet could still login blind. Booted to safe mode, changed the driver back to 174.74, and I’m back in Windows again without the ability to use CUDA once again.

Obviously I’m not going to cry over this, the release notes said it was unsupported, and I just wanted to check it was. I’ll just be patient now for the final release.

Thanks for the follow up. I’ll assume the Device-to-Device number you saw was an artifact of untested/unsupported drivers, and be patient for the final CUDA 2.0 release. :)

I had similar experience with Everquest running on an 8800GTX Vista Business 64bit. I set the NVidia control panel to let the application set the quality (it was set on hightest quality orginally) and restarted. Everquest then ran fine. I am not sure if it was the restart or lower quality rendering that fixed the problem. In other words, your driver hack may not be what caused your game to flake out.

-Mark Granger