Will Edgington writes:
>Or does someone already maintain a page with timings
>for non-Intel CPUs? I don't recall seeing one.
Yes, there was such a page, but the person maintaining it never did much
beyond some initial small timing samples.
George, if you agree that the page is basically defunct, I'll be happy to
maintain it.
I suggest making a table listing both the raw per-iteration timing for a given
code X on a given machine, but also defining a nondimensional parameter that
measures the performance of code X relative to Prime 95, normalizing for clock
speed, e.g. let rp = relative performance:
(code X iter. time)/(code X machine clock rate)
rp = ------------------------------------------------------------------- .
(Prime95 iter. time)/(Pentium clock rate)
Thus, Prime95 = 1 by definition. Any rp < 1 indicates relastively poorer
performance
than Prime95, whereas rp > 1 indicates relatively better performance.
The table should include complete timings for a wide range of FT lengths,
including
non-power-of-two runlengths. Then, codes lacking non-power-of-two runlengths
would
list timings for those based on the timing for the next-higher power of two,
i.e. would
automatically be penalized for lacking, say, N = 384K, by having to use the N
=512K
timing (and rp relative to the Prime95 384K timing) in that column.
We should strive for as wide a range of machines and subtypes as possible, so
as
to quantify more-subtle things such as behavior vs. cache size, which can be
very
different for different implementations (e.g. codes like Prime95 and
MacLucasUnix which are designed for small-cache, bandwidth-limited systems
will perform well
on such systems, but may tend to benefit less from larger caches).
And of course we should allow for plenty of room for comments about the
various
codes, e.g. whether it needs a particular compiler, whether binaries are
available,
whether it includes an integrated factorer, who maintains it, etc.
-Ernst
_________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
