[Date Index] [Thread Index] [Author Index]

Re: CUDA support in 7.0?

• To: mathgroup at smc.vnet.net
• Subject: [mg93920] Re: CUDA support in 7.0?
• From: Jens-Peer Kuska <kuska at informatik.uni-leipzig.de>
• Date: Fri, 28 Nov 2008 05:10:52 -0500 (EST)
• Organization: Uni Leipzig
• References: <ggj7dd$j36$1@smc.vnet.net> <ggjf2b$lun$1@smc.vnet.net> <gglsrj$8ev$1@smc.vnet.net>
• Reply-to: kuska at informatik.uni-leipzig.de

Michael Weyrauch wrote:
> Hello,
> 
>    I do not really understand what you are talking about here, Jens,
> and on what basis.

I' talking about that a general multi-platform system like Mathematica
can't use very special hardware in general, because the very special 
hardware does not exist in general.

As to see from

http://www.gpgpu.org/

"With desktop systems based on Tesla GPUs, Mathematica users will be 
able to perform complex, data-intensive computations right at their 
desk, negating the need to write native C programs or wait for time on a 
public cluster, a process which can often take days or even weeks.

The CUDA accelerated version of Mathematica is expected to be available 
in Q1 2009."

> 
> But I know from a very good talk (on the Mathematica users conference 
> and not under NDA) by one person of Wolfram Research and another one of 
> Nvidia, that they are experimenting seriously with CUDA.

What may be the difference between "experimenting seriously" and 
"support it in a product" ??

> (Alltogether it seems not to be that simple, it was said that one 
> requires good support by the graphics card manufacturer in order to get 
> it going).  

Only *one* Graphics card manufacturer support it at all. AMD/ATI will 
use its Stream Computing/ Brooke+ language 
(http://ati.amd.com/technology/streamcomputing/sdkdwnld.html) and Intels 
Larrabee will also have it's own languange.

Moreover, it seems that BrookGPU /Brooke+ 
(http://www-graphics.stanford.edu/projects/brookgpu/)
will become the standard and not CUDA

Especial because nVidia try to sell its newes graphics cards
for CUDA, while BrookGPU work well even with older ones.

Also, if I understood the speakers well, one cannot go
> beyond the single precision limit in general. 

GTX260, GTX280 can do that and this is so general as to have a GeForce 
Graphics card and not a Intel, ATI, S3 ...

So, I understand that only
> specific numerical code parts can be executed on the graphics card (?).
> 

What else ?

> If such support makes it into a future version of Mathematica, I, of 
> course, don't know.

and it seems that you also "not really understand what you are talking 
about here"

Regards
   Jens

> 
> Michael
> 
> Jens-Peer Kuska schrieb:
>> Hi,
>>
>> a) Mathematica run on Solaris and system with ATI/AMD cards
>>     so there can no general support fo CUDA
>> b) it is simple to make a MathLink program that use CUDA and
>>     *this* is the support that exist since version 2.0
>>
>> Regards
>>    Jens
>>
>> Amir wrote:
>>> We are thinking of making some equipment purchases for CUDA use. Can
>>> someone detail what the CUDA support is in this release of Mathematica
>>> if any? We haven't received a copy of 7.0 yet and the new
>>> documentation doesn't mention CUDA or how it's integrated, but it has
>>> appeared in Nvidia marketing blurbs. For example, will it be use-able
>>> with a only newer Nvidia cards? How does it get around the use of
>>> single precision in current cards?
>>>
>>> Thanks,
>>> Amir.
>>>
>