Latest drivers (NVidia and ATI)

[Jason G]

Nah, It's only prime sizes, so don't get too excited  .  There are many fixes & small speedups over 3.1 though, where 3.1 was not usable for general release but 3.2 looks like it might be.  That will allow some good stuff down the road.

[Raistmer]

And some bad news - analyser (profiler?) will be included only in professional ($349) version of NSight , standart (free) version has only debugger :/

[Fredericx51]

Hi, another  Fund Raise, maybe, if it's really necessary  ?

(Just got my 470 and 480 (GTX) working). On an ASUS P5E Mobo and 850Watt PSU.
A bit of topic.......

[Jason G]

And some bad news - analyser (profiler?) will be included only in professional ($349) version of NSight , standart (free) version has only debugger :/

Nah, cuda visual profiler is part of the toolkit/SDK.  nSight has more advanced stuff like visual studio interopability. 

[msattler]

Any sightings of 3.2 yet?

EDIT.....
Answered my own question....I see the 3.2RC SDK is on the nvidia site now.........

[Vyper]

Any sightings of 3.2 yet?

Well the sighting seem to be in the 260x driver set for sure.

I also recieved an nvidia email that they have released 3.2 for developers now stating speed improvements in CUFFT radix and CUBLAS for fermi architecture.
This was something i read too in the readme amongst deprecated functions:
CUFFT supports batched transforms > 512 MB , The previous version of CUFFT failed when (batch size * transform size * datatype size) for a 1D single-precision transform exceeded 512MB. This has been fixed so that now the total size can be as large as the device memory capacity allows

As we saw earlier we got some speed boosts running latest drivers also with plain Cuda 3.0 code too so i'm about to try the later driver set on my 460 card to see if true workunits benefit it also compared to the quick WUs.

Kind regards Vyper

[perryjay]

Ok, is there any good reason for us non-fermi types to go for this new driver before it comes out for everybody? Do you guys need any testing from us on this or can we just go back to sleep and wait for it?

[Vyper]

Ok, is there any good reason for us non-fermi types to go for this new driver before it comes out for everybody? Do you guys need any testing from us on this or can we just go back to sleep and wait for it?

I don't know yet, the best thing is to run a set of benches before and after upgrading to the newer driver set.
Since i don't want to experiment with my own main host which runs 295's i too have to wait and see what other finds out.

But best thing is to run knabench and compare the timings and post here in the forum so we know..

Regards Vyper

[Jason G]

I'd suggest the WHQL are the way to go for the moment, though there are some benefits in the beta drivers.  When I switch to Cuda 3.2 you'll need to update, but that won't happen until this weekend at the earliest.

[msattler]

Is this going to be as simple as dropping the new dlls into the existing installation with the new driver in place?
If so, somebody post the new dlls and I can give it a go.
I have the new beta driver running on my dual GTX295 and my dual GTX465 rigs.

[Jason G]

Hi Mark,
   Short answer: No.

Long answer:
   Cuda SDKs & Libraries from 3.0 onwards are 'strongly versioned'.  That means apps built with a certain SDK are meant to be used with certain DLLs.  The New SDK with New DLLs will require a new App.
I would not advise playing swippy swappy with DLLs, as was possible in the past.  You could get lucky and it works, but if you break it you own both parts    Also, the 3.2 SDK is so far accessible to registered developers only, and is under NDA, so no-one will be posting Cuda 3.2 DLLs around here while I'm on watch 

[msattler]

Hi Mark,
   Short answer: No.

Long answer:
   Cuda SDKs & Libraries from 3.0 onwards are 'strongly versioned'.  That means apps built with a certain SDK are meant to be used with certain DLLs.  The New SDK with New DLLs will require a new App.
I would not advise playing swippy swappy with DLLs, as was possible in the past.  You could get lucky and it works, but if you break it you own both parts    Also, the 3.2 SDK is so far accessible to registered developers only, and is under NDA, so no-one will be posting Cuda 3.2 DLLs around here while I'm on watch 

Short response: OK, mate.

Long response:  Understood......I did not realize that NDA applied, I would not ask anybody to break that here.
I will hold my kibble until you guys play with it a bit and nvidia makes it freely available.
I was going offa the last experience we had when a simple dll upgrade made lots more crunching.

[Jason G]

LoL,
   Yeah I strongly suspect. it's that DLL swippy-swappying going on all over the place that prompted nVidia to make the changes. I can imagine all the nVidia call-centre help desk people being fooded with 'My Cuda broke' calls because 2.2 & 2.3 weren't 100% functionally identical, but worked for most purposes with 2.3 DLLs under 2.2 or earlier builds with increased performance.

It's pretty hard to diagnose what the problem is if something crashes, when the app was built with some Cuda version, but the libraries are goodness knows what version.

A quick rebuild on the weekend should indicate if something should be put through express testing ready fior 3.2 release.  I'm guessing at this point it might be able to wait until a bit more streamlining & fine tuning is done, as I'm not expecting a huge difference,, but never know, I've been surprised before.

[msattler]

Best of luck with it......
You can PM me if you want me to give it a go.

Meow meow.

[perryjay]

Yeah, what Mark said!  You know me, always running around here in a state of mass confusion but willing to try!