Elmer, <..provide the data that proves you wrong.>
There is nothing in this world that can prove me wrong
on this subject. There might be only misunderstanding
(or unwilling to learn) on your part, or my limited
ability to explain the point. I will try again.
Your mistake is that are using an inappropriate example.
First, we are talking here about the mass desktop market,
where typical applications are Word, spreadsheets, games,
scheduler, net browsing/email, easy-CAD home design, etc.
These applications have certain "affinity" of data they work
with, so the desktop systems are designed with appropriate
size of L2 cache that fits these data sets and is economical
enough.
For this typical application workload, the system cache
size determines how many memory references will miss
that cache. Only these misses contribute to the time
wasted in access to external DRAM and determine the
difference between 66MHz and 100MHz DRAM/SDARM timing.
Now, you picked up the SPECfp95 benchmark. As you might
be aware, this benchmark uses examples of most challenging
scientific calculations like shallow water and
Navier-Stokes equations, turbulence modeling and weather
prediction, quantum chemistry and Maxwell's equations.
Those are hardly the applications people are running at
home. One of the problem with the SPECfp apps is
that they require either huge memory that does not
fit any reasonable cache, or their data sets have very
poor address locality. All this results in excessive
L2 misses that must go to main memory and stall CPU
for many bus clocks. And since only this portion of
run-time is affected by changes in the bus speed, you
see some speedup.
Actually, why don't you pick SPECint? Winstone98?
BAPCO sysmark?
In conclusion, the desktop mass-market computers are
not designed for applications from the SPECfp95 suite,
and inappropriate use of a PC does not prove me wrong.
What do you think, if you decide to hit a nail, which
device will be better, Xeon or K6?
I bet AMD will lose :-) |
