[Beowulf] 3.79 TFlops sp, 0.95 TFlops dp, 264 TByte/s, 3 GByte, 198 W @ 500 EUR

Vincent Diepeveen diep at xs4all.nl
Thu Dec 22 08:04:09 PST 2011


As for HPC, do they need to do that - did AMD already release a  
driver for example for OpenCL for the HD 6990 that's using BOTH gpu's?

I had back then bought directly a HD 6970 card.

Once the driver for the HD 6970 was there for linux,
we were months further and the price of the HD 6970 had dropped  
considerable again at the shops.

Multiplying 32 x 32 bits is slow at AMD gpu's, as it needs all 4  
procesing elements for that. Nvidia wins it bigtime there.

Fast at AMD seemingly is 24 x 24 bits, yet of course you also need  
the top 16 bits of such multiplication.
Then after a while i figured out that OpenCL has no function call for  
the crucial top 16 bits.

Initially there was a poster on the forum saying that this top 16  
bits was casted onto the 32 x 32 bits anyway, so would be slow anyway.

Raising a ticket at AMD then, we speak again about months later,  
revealed that the hardware instruction i found in their manual that's
doing the top16 bits of a 24x24 bits integer multiplication, total  
crucial for factorisation work, that this indeed runs at full throttle.

Some AMD engineer offered to include it, i gladly accepted that, of  
course we were months later by then.
We are 1 year further nearly now and it's still not there.

This HD6970 so far was a massive waste of my money. Can i ask my  
money back?

You sure this will go better with HD7970 not to mention the soon to  
be released HD7990?

 From HPC viewpoint AMD has a major software support problem so far...

...also i noticed that the problem was not so much being 'busy', as i  
saw relative few tickets got raised for their gpgpu team.

Regards,
Vincent

On Dec 22, 2011, at 3:57 PM, Eugen Leitl wrote:

> On Thu, Dec 22, 2011 at 09:43:55AM -0500, Prentice Bisbal wrote:
>
>> Or if your German is rusty:
>>
>> http://www.zdnet.com/blog/computers/amd-radeon-hd-7970-graphics- 
>> card-launched-benchmarked-fastest-single-gpu-board-available/7204
>
> Wonder what kind of response will be forthcoming from nVidia,
> given developments like http://www.theregister.co.uk/2011/11/14/ 
> arm_gpu_nvidia_supercomputer/
>
> It does seem that x86 is dead, despite good Bulldozer performance
> in Interlagos
>
> http://www.heise.de/newsticker/meldung/AMDs-Serverprozessoren-mit- 
> Bulldozer-Architektur-legen-los-1378230.html
>
> (engage dekrautizer of your choice).
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin  
> Computing
> To change your subscription (digest mode or unsubscribe) visit  
> http://www.beowulf.org/mailman/listinfo/beowulf




More information about the Beowulf mailing list