Latest drivers (NVidia and ATI)

[Raistmer]

released CUDA 3.0beta time to recompile

Prev CUDAs didn't increase speed much from recompile. Most speedup came just for using new CUFFT library with older build.
That is, if we will have ability to get CUFFT 3.0 some speedup could be recived "for free" I hope

[Claggy]

released CUDA 3.0beta time to recompile

Prev CUDAs didn't increase speed much from recompile. Most speedup came just for using new CUFFT library with older build.
That is, if we will have ability to get CUFFT 3.0 some speedup could be recived "for free" I hope

You do know that CUDA Toolkit 3.0 beta released for registered developers?

Claggy

[Raistmer]

released CUDA 3.0beta time to recompile

Prev CUDAs didn't increase speed much from recompile. Most speedup came just for using new CUFFT library with older build.
That is, if we will have ability to get CUFFT 3.0 some speedup could be recived "for free" I hope

You do know that CUDA Toolkit 3.0 beta released for registered developers?

Claggy

nVidia has too "corporation-oriented" registration form for beta program...

[Jason G]

nVidia has too "corporation-oriented" registration form for beta program...

Nah, you just tell the truth on the form, in the most optimistic way possible (Which strategy matches with nVidia's marketing strategy quite well  , think keywords like signal processing, optimisation, radar & deep space ).  They know who we are (I reckon).  Will be taking a look when I get the time (since already registered for a fair while), and see if there's any useful improvements, or if it's just a major revision for as yet unreleased hardware compute capability. (Which I feel is likely) .. nothing there last week when I looked, but will look again.

[Devaster]

released CUDA 3.0beta time to recompile

Prev CUDAs didn't increase speed much from recompile. Most speedup came just for using new CUFFT library with older build.
That is, if we will have ability to get CUFFT 3.0 some speedup could be recived "for free" I hope

You do know that CUDA Toolkit 3.0 beta released for registered developers?

Claggy

nVidia has too "corporation-oriented" registration form for beta program...

i am not in corporation and i am there ... nobody was verifified my given infos ....

[Raistmer]

i am not in corporation and i am there ... nobody was verifified my given infos ....

Ok, will try to fill that huge form again 

[Pepi]

Cuda 3.0 is already on some of computers (look at thread)  BOINC startup CUDA info

GeForce 8700M GT (driver version 19539, CUDA version 3000, compute capability 1.1, 512MB, 80 GFLOPS peak)

[Jason G]

Reviewing the Beta download information, there are no obvious performance related improvements related to 'existing  builds & hardware' that currently run with 2.3 DLLs  (unlike the 2.2-> 2.3 transition which featured much faster FFT processing), however that doesn't mean there couldn't be *some*.

The Cuda 3.0 Beta SDK's are under the usual registered developer NDA's so that means DLL's etc cannot be publicly released until nVidia say so ... And may not hold any benefits 'till you buy some as yet unreleased card anyway....

Put another way, the primary improvements in Cuda 3.0 appear to be DirectX and OpenGL/OpenCL interoperation, and preliminary introduction of Fermi  feature support, rather than anything useful to us right now...

Experimental builds will evolve in the direction that needs that stuff, but can only go for public test / release when such hardware & builds exist ...  and such hardware I imagine is likely to be well out of my price range.  So People:  please stop PM'ing me about Cuda 3.0 builds .. you're not getting a build until it (a)would be of benefit to you AND (b) nvidia says so. (Probably @Fermi release later in the month), and possibly (c) Affordable GT300 based cards are released (And I'm able to get one).

Aside: Also stop asking ( Me by PM here and at main) about 64 bit builds please ... Current (V12)releases run on the GPU entirely, so the OS/library doesn't matter a toss ...   V13 Hybrid Is a different story .. but is very much Pre-Alpha, so 64 bit is ignored for development consolidation purposes.

Regards, Jason

[Claggy]

Thanks Jason for the clarification,

Claggy

[Devaster]

so for now replacing libraries doesnt work i must recompile ....

[Vyper]

Yes

That's true it seems, i've tried and it doesn't work unfortunately.
But i also think that Cuda 3.0 is more or less directed towards upcoming Fermi architecture cards and i suspect that if you run Cuda 3.0 with non fermi hardware it will not speedup so much more than 2.3 can deliver.
All the advantages with 3.0 will be directed towards Fermi cards so nVidia probably can sell more hardware.

Regards Vyper

[sunu]

Nvidia has stated numerous times that cuda is not only a software thing (just extensions to a programming language) but a combination of software-hardware. And since Fermi is a new architecture bearing a new cuda version (3.0), it is more than likely that previous hardware won't benefit much if anything at all.

[Vyper]

I found this on another site..

"Return values are as for all other cufftPlanXxx functions. Thus to plan
   a batch of 1000, 2D, double-precision, complex-to-complex transforms of
   size (128, 256), you would do:

       cufftHandle *myplan;
       cufftPlanMany(myplan, 2, { 128, 256 }, NULL, 1, 0, NULL, 1, 0, CUFFT_Z2Z, 1000);

   Note that for CUFFT 3.0, the layout of batched data must be side-by-side
   and not interleaved. The inembed, istride, idist, onembed, ostride and
   odist parameters are for enabling data windowing and interleaving in a
   future version."

A recompile and correction of the code seems to make the FFT part working in Cuda 3.0..

Kind regards Vyper

[Jason G]

Cheers for the warning... be some time before we need that, but handy.  (While Cuda 3.0 SDK is in registered developer status, it's indeed offering little worth to us until new hardware is released )

[Devaster]

it is for free download for now ....