The solution to the reported oddity is to remember that the CPU speed is
only part of the overall system performance. While CPU speed has jumped
upward considerably (say from 200 Mhz to 1000 Mhz), the basic SDRAM memory
speed is still only 100 Mhz, even on the Compaq 7998. (Yes I know that
there is 133 Mhz memory, etc. I am not openning that issue right now).
The basic memory speed limitation is forcing the Level 2 cache to play a
more important roll in keeping the CPU busy. When you calculation fits in
L2 cache, you run fast, when it starts to fall out, you run slower due to
memory bandwidth issues.
While alot depends on how efficiently the cache is utilized, let me give you
some examples to illustrate this point. This does not relate directly to
prime95, but should illustrate the point. The software used is Proth, by
Yves Gallot. The base speed for this is a 200 Mhz Pentium MMX. That scores
about 100. The others are relative to this. These are actual machine
measurements.
Pentium MMX Pentium 3 Pentium 3 Pentium 3 Athlon
Number Size 200 Mhz 450Mhz 700Mhz ?L2 700Mhz 512L2 1Ghz 512L2
3*2^64000 + 1 114% 334% 583% 630% 1072%
3*2^128000 + 1 114% 347% 563% 638% 842%
3*2^512000 + 1 118% 357% 268% 464% 735%
3*2^1024000 + 1 126% 341% 263% 452% 631%
3*2^4096000 + 1 128% 316% 272% 462% 442%
Notice at smaller number sizes, the productivity relates to CPU speed
closely, but at larger number sizes, the productivity degenerates more for
the higher CPU speeds. This Athlon measured is the Compaq 7998.
The big jump between the Pentium MMX and the Pentium 3 is the addition of a
second FPU unit starting the the Pentium Pro line. That means Mr. Gallot as
done a good job of programming to utilize the both FPUs efficiently.
Lastly notice the cheap 700 Mhz machine in the middle which degenerates to
slower than my 450 Mhz machine. That has a lousy memory interface design.
Just because the box says 700 Mhz P3 does not mean the rest of the design is
any good.
At 03:20 AM 7/7/00 -0700, you wrote:
>Hi All:
>
>On several Athlon machines, from 750MHz to 1000MHz, P95 takes
>about the same number of clocks per iteration doing a factoring assignment,
>in the present case 4.75 billion. The Compaq 7998 settles at 5.28 billion
>for the same task, but can be driven down to 4.75 briefly by re-booting,
>running ReCache, and then some start/stop/start shenanigans. Soon the number of
>clocks starts to drift upward, and settles at 5.28/5.3 again. Other trial
>factoring stages show the same anomaly: drift-and-settle always at a higher
>than expected number. Trying the Advanced/Time test on the default 10
>mill exponent, I notice unusual variation: from .125 (expected) to .134
>(dreadful) for a 1GHz machine.
>
>Any insights?
______________________________________________________
_/_/_/ David Underbakke [EMAIL PROTECTED] _/_/_/
_/_/_/ Aramit Technologies,Inc. [EMAIL PROTECTED] _/_/_/
_/_/_/ Res:(763) 577-1400 Bus:(763) 577-1401 _/_/_/
_/_/_/ Fax: (763) 577-1414 _/_/_/
_________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
