Prev: A link to news on help of tapping power of GPU
Next: Prime Number Fluid model data and graphs
From: Clark Smith on 20 May 2010 14:10
According to this link

http://www.phoronix.com/scan.php?page=article&item=intel_corei3_530&num=6

one can compute some 148 signatures per second using 4096-bit RSA moduli
on an AMD Phenom II X3 710 CPU. Is this performance per core, or is it an
aggregate of such operation running in parallel on each of the four cores
in this CPU?
From: Mok-Kong Shen on 20 May 2010 15:40
Clark Smith wrote:
Is this performance per core, or ......

I am ignorant, but to my knowledge benchmarking is as a rule done on
a computer when the application chosen for the benchmark is running
without any concurrent applications.

M. K. Shen
From: Datesfat Chicks on 20 May 2010 17:12
"Mok-Kong Shen" <mok-kong.shen(a)t-online.de> wrote in message
news:ht436j$n22$00$1(a)news.t-online.com...
> Clark Smith wrote:
> Is this performance per core, or ......
>
> I am ignorant, but to my knowledge benchmarking is as a rule done on
> a computer when the application chosen for the benchmark is running
> without any concurrent applications.
>
> M. K. Shen

The equation is more complicated than that nowadays. Most CPU's for the
personal computer market are multi-core. You really have to count that as a
standard computer nowadays, even with no "concurrent applications".

The background is that transistor switching speed is limited by laws of
physics (and can't follow Moore's law), but the number of transistors you
can pack on a chip does follow Moore's law. So, you can now put zillions
and zillions of transistors on a chip, but none of them can switch that much
faster than their ancestors five years ago ... how do you use that?

A lot of the design features to use the extra fab capability are at the CPU
level (pipelining, caching, various speed vs. size tradeoffs), but once you
get beyond that and you still have space left over, the natural idea is to
put more than one CPU on a chip.

http://en.wikipedia.org/wiki/Multi-core_processor

That is a personal computer nowadays. That IS the platform.

Whether concurrent applications are running is a separate issue.

My assumption is that the benchmark is for the thing as a whole, i.e. not
"per core".

Datesfat

From: Mok-Kong Shen on 20 May 2010 18:48
Datesfat Chicks wrote:

> Whether concurrent applications are running is a separate issue.
>
> My assumption is that the benchmark is for the thing as a whole, i.e.
> not "per core".

If there are multicores the performance would anyway also depend on the
performance of the compiler in well exploiting the available resources,
I suppose. If there were concurrent jobs running, then the benchmark
result obtained would be dependent on these. Hence benchmarking should
be done with a single job running, unless one is investigating the
proformance of a computer under a general multitasking scenario, which
of course also has its own sense, in which case the job mix needs
however to be properly specified. Thus in OP's case the figure in
question is surely for the whole processor as you said above.

M. K. Shen


From: Maaartin on 20 May 2010 20:30
There are three cores in AMD Phenom II X3 710 CPU, not four.

On May 21, 12:48 am, Mok-Kong Shen <mok-kong.s...(a)t-online.de> wrote:
> Datesfat Chicks wrote:
> > Whether concurrent applications are running is a separate issue.
>
> > My assumption is that the benchmark is for the thing as a whole, i.e.
> > not "per core".
>
> If there are multicores the performance would anyway also depend on the
> performance of the compiler in well exploiting the available resources,
> I suppose. If there were concurrent jobs running, then the benchmark
> result obtained would be dependent on these. Hence benchmarking should
> be done with a single job running, unless one is investigating the
> proformance of a computer under a general multitasking scenario, which
> of course also has its own sense, in which case the job mix needs
> however to be properly specified. Thus in OP's case the figure in
> question is surely for the whole processor as you said above.

The most efficient way for computing this benchmark is probably to
start just one thread per core. This way there's no need for
parallelization of a single job and no unnecessary pressure on the
cache (although here it's no problem anyway, since even the L1 cache
is large enough). I assume, the reason for AMD winning over Intel in
this benchmark is having one more core (not counting virtual cores due
to hyperthreading).
 |  Next  |  Last
Pages: 1 2 3
Prev: A link to news on help of tapping power of GPU
Next: Prime Number Fluid model data and graphs