Re: implemented features (Re: Decision time....)

[Rahul Siddharthan <rsidd120@xxxxxxxxx>]

"km b" <kmb810@gmail.com> said:
>On 6/6/07, Erik Wikström <erik-wikstrom@telia.com> wrote:
>
>> The place where he found the 50 % performance increase is indicated
>> right in the text you quoted, in *his* binaries compiled from *his*
>> code. He also did mention that this seemed to be true for *some* parts
>> of the GSL, not all of it. So unless you can get hold of the same
>> program as he is talking about and showing that there's no difference
>> between 64-bit and 32-bit machines then there's nothing you can do to
>> prove him wrong.
>
>Rahul is a theoritical physicist and I am asking him to compare
>oranges vs oranges and not apple vs oranges to back his claims.

My benchmarks were on the same machine.  Not the same "platform", the
same *machine*.

In your case, you're giving different architecture arguments (prescott
vs nocomo) and yet claiming it's apples-to-apples.

You don't say what compilers you were using on the two platforms, or
even whether they were the same (gcc4 can sometimes be significantly
faster).

(As an aside, you don't know the difference between PDF and
postscript.)

All that apart, you still see a 25% increase in peak mflops in the
64-bit graph.  I don't see why you don't consider that significant.

Finally, you are testing on Intel and I am testing on an AMD Opteron.
It is widely believed that opterons do much better in 64-bit mode
(unfortunately, nearly all the published benchmarks are 32-bit and
basically useless).  Cray, Sun and many others in high-performance
computing have chosen opterons for that reason.

Rahul