| >>>>> "VVS" == Vadim V Sytnikov writes:
VVS> Gregory Boduch wrote:
>> As a side note, there seems to be a problem with the Pentium Pro cpu. A
>> Linux benchmark called Oscillates, which simply tries to send as many
>> wavetable-lookup oscillators as it can to the dsp, returns a score of 42
>> oscillators on a P200 w/ 32Meg, 36 oscilators on a P166 MMX, and only
>> 28 on a PPro 200 w/ 32Meg and 27 on a dual PPro 180 w/ 32Meg.
>>
VVS> If you REALLY want to compare several Intel CPUs with some benchmark,
VVS> you should do the following:
VVS> 1) install some 32-bit Microsoft OS (I'm very sorry, but you
VVS> CANNOT neglect this -- see below),
VVS> 2) use Intel Reference C Compiler (aka Proton) to compile your
VVS> benchmark with proper "target optimizations" flags for each
VVS> of your CPUs.
VVS> You cannot use non-MS or non-32bit OS since Proton does not support them.
VVS> You should not use compiler other then Proton since in such a case
VVS> chances are your compiler will generate more "blended" (Intel's term)
VVS> code then necessary -- i.e. insufficiently tuned.
VVS> It doesn't mean that Proton is the very best: it simply means that
VVS> it should be used for benchmarking. I use Symantec C for the work
VVS> but Proton for benchmarks and selective optimizing.
Hey! Hallo there!
Your benchmarks will not say that much if you use _diffrent_ compilers
for work and for benchmarks. In order for them to say something you
must use the same compiler and a benchmark which is similar enougth to
an real application. You can benchmark a lot of diffrent things, but
creating a diffrence in _both_ application and compiler will make the
benchmark numbers say _less_ about preformance. That would be equalent
of showing a care with turbo drive and testing it on a race track
while trying to sell an care to a customer which is going to bye an
standard (non turbo) care for city trafic only. Sure, the race car
version is nice and has certainly nice acceleration, but it doesn't
say much about the car in city trafic. My analogy has more preformance
difference, but that problem is about the same. Showing the turbo
version car in city trafic would be much easier to compare, since you
now can more easilly see where the turbo would make the diffrence or not.
VVS> Next problem -- good *CPU* benchmark must NOT contain any I/O
VVS> calls or library function calls WITHIN the test loop. Oscillates
VVS> obviously does not follow this guideline.
VVS> P6 is by no means worse -- but it is MUCH more sensitive to all this.
VVS> And in general, we should compare computers, not CPUs.
In this case we where even comparing or trying to get a measurement of
how fast some computer would deal with some csound processing. I would
rather recommend that a suite of csound processing scores and
orchestras would be created to measure that instead.
Magnus
|