Future release improvements What are you planning?

Hi everybody, I’m writing my thesis about CUDA researches and, if possible, I wish to know something about planned improvements that will be included in the next CUDA toolkit versions.


3D textures will be supported in next release, probably. There is also a longer list by an NVidia guy (I think in the 1.1 release notes), you might find it with a forum search.

CUDA support for Mac OS X was demonstrated at SC2007, but hasn’t been released yet. Presumably Vista support is also going to available “real soon now.” wumpus mentioned 3D textures, which are a frequently requested feature. Double precision support will probably happen sometime in 2008, though that really is a hardware change.

Other than that, the Nvidia employees generally keep pretty quiet about future features.

wumpus and seibert have pretty much nailed it :)

Thank you.
Your answers are very helpful.

Double precision floating point is supposedly due late 2007… wait… its 2008 now? Hm.

I’m curious if the SDK is just lagging behind the hardware at this point. The G92 core is out, but there dont appear to be any new CUDA-specific features for it… yet? ;)

G92 has atomic functionality & async memcopy (during kernel execution) This is supported by CUDA 1.1 so no lagging behind.

There were requests (also from wumpus) that the “PureVideo” engine in the G92 be exposed to CUDA, since that sounds like it can do some useful operations. We still haven’t got access to it, and it isn’t clear if we ever will. Either its low on the priority list, or some kind of voodoo licensing is involved. Almost anything involving video codecs tends to be messy.

Another feature that’s supposedly coming is accelerated copy between multiple GPUs. The API is exposed in CUDA 1.1 but direct copying is not used yet in that version (as it would need a big change in the memory pinning api of the driver, I believe)

Yup, that’s that’s one of the reasons I’m using (and working on) the Dirac codec, it’s free of those issues. Of course, it’d be ideal if the VP(2) is exposed and turns out to be programmable enough to support a custom codec like that.

could we have an official estimation release date of the double precision floating point

( Q1 , Q2, Q3, Q4, 2009,…)

NVIDIA employees are not allowed to comment on release dates, but Google is your friend :)

According to this, CUDA 2.0 (with double precision support) is scheduled for Q1 2008. So, it may be Q1 or Q2 (since presentation is pretty old, but all prior checkpoints were reached in time).

Hmm, I missed the visual HW debugger when it came out. You have the link for me? :D

Um, I think I’ve confused it with Visual Profiler, sorry :(

Well, actually at this time I am making no errors ™ so I have now more use for the profiler, but there was a period 2-3 months ago zhen I would have ‘killed’ for a visual debugger.

I was wondering about this too. Do the new 9000-series of cards support it?

No, there is no released hardware with double precision support.