FYI: Current SPECfp landscape...

[Josip Loncaric]

Greg Lindahl wrote:
> 
> On Thu, May 10, 2001 at 01:11:39PM -0400, Josip Loncaric wrote:
> 
> > Per GHz, the F77 P4 "win" factors are 1.25-1.59, while the F90 P4
> > "loss" factor is 1.38.
> 
> What's the point of per GHz comparison? The P4 and the Athlon aren't
> available at the same GHz, and the wide memory system of the P4 is a
> large part of its advantage.

The point is data compression: I can convey the first order
approximation in one number, with minor (<5%) errors.  Also, I can
estimate the relative performance in a fairly consistent manner and
conclude that on SPECfp2000 an X GHz P4 would be roughly equivalent to a
1.1*X GHz Athlon.

You do a linear fit near some operating point, then use a linear model
to get approximate numbers near that operating point.  It works
reasonably well.  Performance does node scale exactly with GHz, but the
individual SPECfp2000 benchmark numbers are not too far from what you'd
expec on the GHz basis.  Also, "per GHz" comparisons tell you something
about a particular architecture.  The Athlon FPU is expected to
outperform the P3 FPU by 4:3 per GHz, but on SPECfp2000 P4 benchmarks,
this effect is *not* obvious.

Here are Intel's SPECfp2000 numbers (D850GB motherboard, PC800 RDRAM):

P4 linear fit: SPfp2K ~ 374 * [P4 GHz]
P4 GHz:        SPfp2K:   predicted by GHz:
1.3               511    486 (-4.8% error)
1.4               538	 524 (-1.7% error)
1.5               558    561 (+0.6% error)
1.7               608    636 (+4.6% error)

Here are the Athlon numbers (Gigabyte GA-7DX motherboard, PC 2100 DDR
SDRAM):

Athlon linear fit: SPfp2K ~ 340 * [Athlon GHz]
Athlon GHz:    SPfp2K:   predicted by GHz:
1.2               417    408 (-2.0% error)
1.33              445    453 (+1.7% error)




> 
> > We'd like to test our own code compiled with our own compilers.  The
> > overall 2:1 performance uncertainty is too large to ignore.
> 
> Since both Intel and PGI make their compilers temporarily available
> for free -- beta-test until September for Intel, trial version from
> PGI -- you might as well test with the best compilers, instead of the
> ones you happen to own.

We own PGI compilers.  Intel compiler availability is *not* the
problem.  Our problem is that we'd like to test hardware we do not yet
have in order to determine if we should have it.  This is a Catch 22
situation...

Sincerely,
Josip

-- 
Dr. Josip Loncaric, Research Fellow               mailto:josip at icase.edu
ICASE, Mail Stop 132C           PGP key at http://www.icase.edu./~josip/
NASA Langley Research Center             mailto:j.loncaric at larc.nasa.gov
Hampton, VA 23681-2199, USA    Tel. +1 757 864-2192  Fax +1 757 864-6134