dfox wrote on Sun, Jul 07, 2002 at 10:53:12AM -0700 :
> Well, looking at the burnK7 (or other portions of cpuburn) I was 
> very skeptical as to how it could really stress the system, as it
> ks just a couple of floating point instructions in an endless
> loop. But from others' input, it seems to peg the cpu temperature 
> meter pretty high, higher than i would think for such a piece of
> software. The loop is too short to not fit in cache (or probably
> the instruction pipeline itself -- that depends on the size, I'm not
> sure) and depending on CPU timing issues and stalls (usually affecting
> older Pentiums) it might not exercise the processor as much as is
> desireable.

To me this means that it *will* exercise the processor.  Don't forget
about cache speed.  L1 cache is much faster than L2 (external) cpu
cache.  If the entire program fits inside cache and never has to insert
the wait states for external cache to reply, it can run faster, which
should be roughly equivalent to more heat generation (ie flipping state
generates xxx microcalories of heat, flipping state more often generates
more heat).

The second option that you mentioned I had not even considered.  That it
can fit inside the pipeline means it should be able to run even faster
(and parallelization means even more heat generation).  But I would
expect it to be designed such that no pipeline ever finishes before the
previous one.  They can all finish staged in the same ordered they
started, but not out of order.  And it also seems like you would want
the program to be larger than the size of the pipelines so that you
never have idle states waiting for the result.

Interesting thoughts, I like your line of thinking.

Blue skies...           Todd
-- 
  Todd Lyons -- MandrakeSoft, Inc.   http://www.mandrakesoft.com/
UNIX was not designed to stop you from doing stupid things, because 
  that would also stop you from doing clever things. -- Doug Gwyn
   Cooker Version mandrake-release-8.3-0.2mdk Kernel 2.4.18-21mdk

Attachment: msg56124/pgp00000.pgp
Description: PGP signature

Reply via email to