GMP+MPFR to GPU?

[Marcelo E. Kaihara]

Dear Torbjorn,

Thank you for your feedback.

The purpose of showing those numbers on the slide was not to compare GPU 
performance
with GMP, but to evaluate the possibility of using GPUs as cryptographic 
accelerators.

By the time I made the slides, I didn't have access to the fastest GPUs 
and did the measurements on a GTX465
(which was the cheapest Fermi card). Latest high-end GPU models come 
with more cores and higher clock frequencies.

I'll take your remarks into consideration for future evaluation. Thanks 
a lot.

Best regards,
Marcelo.


On 12/08/2010 06:35 PM, Torbjorn Granlund wrote:
> Paul.Zimmermann at loria.fr writes:
>
>    after discussions with Marcelo Kaihara (thanks to him) I was able to reproduce
>    his results with openssl-1.0.0c on a 2-proc 3Ghz AMD Phenom(tm) II X2 B55:
>
>    I configured OpenSSL with GMP as follows:
>
>    $ ./Configure linux-generic64 enable-gmp -lgmp -L$GMP/lib  -I$GMP/include
>
> Should I interpret the command line as if this is for GMP 5.0?  What GMP
> configured for the host processor of this system?  If it is, OpenSSL is
> poorer than I had expected...
>
> My point still holds: the numbers of the slides do not currectly show
> the performance of GMP on RSA-2048.  At best, the numbers are a good
> comparison between OpenSSL's max performance on RSA-2048 on a 1.5 year
> old midrange CPU, with raw code running on the very latest highend GPU.
>
> The main GMP development system. a 3.2 GHz X6 system, gives over 6400
> RSA-2048 decrypt operations per second.  That's without adding the
> OpenSSL overhead.
>


-- 
Marcelo E. Kaihara
EPFL-IC-IIF-LACAL
INJ 333 Station 14
CH-1015 Lausanne
Switzerland
email: marcelo.kaihara at epfl.ch
web: http://people.epfl.ch/marcelo.kaihara
Tel: +41 (0)21 693 64 69
Fax: +41 (0)21 693 64 80