I was mislead by Coppermine FPU benchmarks:
Sorry if I am the last one to learn this.
I had thought that the Coppermine was a major improvement for
Pentium FPU speed for typical workstation use. However, I have learned that FPUmarks are not representative of real workstation use in that they emphasize odd slow instructions like square root.
Much of the floating point work I am involved in (and many others) is primarily multiply and add with almost not divides.
The benchmarks below show the outstanding advantage of of 700 Althon vs the 733 Pentium III (Coppermine).
aceshardware.com
Enter Flops. Flops, programmed by Al Aburto (aburto@nosc.mil), is a very floating point intensive benchmark. The most frequently used instructions are FADD (Floating point addition), FSUB (Subtraction) and FMUL (Multiplication).
This table should make you understand what each Flops test does:
FLops bench FADD FSUB FMUL FDIV
Flops (1) 40.4% 23.1% 26.9% 9.6%
Flops (2) 38.2% 9.2% 43.4% 9.2%
Flops (3) 42.9% 3.4% 50.7% 3.4%
Flops (4) 42.9% 2.2% 54.9% 0%
And here are the results:
K6-2 400 K6-III 400 PIII-600 BX PIII EB 733 Athlon-700
Flops (1) 54,4 54,5 182,7 223 312,6
Flops (2) 58,7 58,8 143,0 174,5 253,6
Flops (3) 86,0 86,1 213,2 260,1 340,3
Flops (4) 118,5 118,6 273,2 332,9 407,7
Nope, the FPU has remained unchanged: a Coppermine 733 is not faster than a PIII 733 would be. Thanks to Flops we finally can see that the triple out of order Athlon FPU is far superior to any other x86 FPU.
In FPU intensive applications with relative small data sets, the Athlon remains the FPU king. |
