VIA C3

Damian Gerow damian at sentex.net
Wed Jun 18 14:44:36 PDT 2003


(Talking about CPUTYPE for Via C3 chips, and whether i586/mmx or k6-3 is
better.)

Thus spake Damian Gerow (damian at sentex.net) [18/06/03 11:31]:
> > Back to the performance-discussion between cputype 586/mmx and k6-3
> > optimization: do you have a suggestion how to benchmark it?
> 
> Well, there's always /usr/ports/benchmarks.  nbench might be what you're
> looking for -- compile it once using 586/mmx support, and once using k6-3
> support.

FWIW, I ran some quick nbench tests on my system.  World and kernel compiled
with i586/mmx.

Here's the results from compiling with k6-3:

    TEST                : Iterations/sec.  : Old Index   : New Index
                        :                  : Pentium 90* : AMD K6/233*
    --------------------:------------------:-------------:------------
    NUMERIC SORT        :          213.91  :       5.49  :       1.80
    STRING SORT         :          17.112  :       7.65  :       1.18
    BITFIELD            :       4.347e+07  :       7.46  :       1.56
    FP EMULATION        :          14.949  :       7.17  :       1.66
    FOURIER             :          2315.8  :       2.63  :       1.48
    ASSIGNMENT          :          3.8335  :      14.59  :       3.78
    IDEA                :          362.06  :       5.54  :       1.64
    HUFFMAN             :          226.37  :       6.28  :       2.00
    NEURAL NET          :          3.1664  :       5.09  :       2.14
    LU DECOMPOSITION    :          164.11  :       8.50  :       6.14

And with i586/mmx:

    TEST                : Iterations/sec.  : Old Index   : New Index
                        :                  : Pentium 90* : AMD K6/233*
    --------------------:------------------:-------------:------------
    NUMERIC SORT        :          220.27  :       5.65  :       1.86
    STRING SORT         :          16.913  :       7.56  :       1.17
    BITFIELD            :      4.3313e+07  :       7.43  :       1.55
    FP EMULATION        :          15.028  :       7.21  :       1.66
    FOURIER             :          2310.2  :       2.63  :       1.48
    ASSIGNMENT          :          3.3773  :      12.85  :       3.33
    IDEA                :          367.09  :       5.61  :       1.67
    HUFFMAN             :          212.37  :       5.89  :       1.88
    NEURAL NET          :          3.0659  :       4.93  :       2.07
    LU DECOMPOSITION    :          155.45  :       8.05  :       5.82

So it looks like it's a tossup between the two, with k6-3 perhaps pulling
out marginally ahead.

Multiple runs of the tests with each flags show numbers extremely close to
the ones posted, if not exactly the same.


More information about the freebsd-hardware mailing list