On Thursday 14 September 2006 21:43, George Woltman wrote:
> >Tony Reix tried to reply but was blocked.  His reply follows:
> >
> >It would be sad to have Prime95 usable only on small
> >multi-CPUs machines. In 5 years from now, probably we will buy
> >16 cores machines for less than 1000 $.

Hmm, the way I see it the extra power is going into the graphics card. CPUs 
seem to be stagnating to some extent, the point being that they're already 
more than powerful enough for most consumer application demands whilst power 
consumption (i.e. heat output, and the associated noise from coolers) is 
becoming increasingly important in a domestic setting.

Graphics is inherently parallel in its nature, whilst the power required by 
ever more hungry 3D applications (games, even the Vista desktop), eye candy 
etc. seems to continue to increase exponentially.

Proprietary application licences will need to evolve before consumers will 
accept multicore systems. Try buying legal copies of e.g. Photoshop CS for  
Windows on a 8 or 16 core system and you'll see what I mean.
> >
> >In fact, all SMP machines are NUMA based, and memory banks are
> >often placed close to some cores/CPUs and thus far or very far
> >from other cores/CPUs. This is called the NUMA factor (Non
> >Uniform Memory Access).

Except for those that share memory controllers!
> >
> >On actual dual-core PCs, probably there is no NUMA factor.

Yes, the memory controller is shared - just as it is on my old Supermicro dual 
PIII mobo.

> >But on 4-cores machines built with recent Intel Woodcrest
> >CPUs, I've seen that the performance decreases when 4 Prime95
> >program were running compared to only one Prime95. Also,
> >performance decrease was higher and higher with larger FFTs.
> >Colleagues saw same kind of scalability problem with other HPC
> >C or Fortran programs.

Gross memory bandwidth must be a factor!

Blocking by incomplete threads must also be a factor of some sort, becoming 
increasingly important as the number of threads increases. However this 
factor disappears if one runs N monothreaded LL tests on a system with at 
least N cores.

Regards
Brian Beesley
_______________________________________________
Prime mailing list
[email protected]
http://hogranch.com/mailman/listinfo/prime

Reply via email to