64 bit performance speedup over 32 bit

[Lars Schouw]

All,

Does anyone know how much performance speedup I can expect by using 64
bit C++ / Windows XP 64 bit over the 32 bit versions?

Did anyone test this under Visual Studio 2005 or Intel C++ 8.1/9.0 ?

My application domain is montecarlo simulations but any test would be
great.

Regards
Lars Schouw

 

[Victor Bazarov]

Does anyone know how much performance speedup I can expect by using 64
bit C++ / Windows XP 64 bit over the 32 bit versions?

How does one compare a truck with a compact car? One can do 70 mph
with a tons in the back across a muddy terrain and the other can do
45 mpg with two people on board. No, I am not pulling any analogy
between cars, trucks, and systems. Just between comparisons.

In any case, have you tried looking on the Web for some performance
comparisons (benchmarks)? Try www.tomshardware.com, for example.

Did anyone test this under Visual Studio 2005 or Intel C++ 8.1/9.0 ?

I bet somebody has.

My application domain is montecarlo simulations but any test would be
great.

You should probably ask this in a Windows programming newsgroup. Or
better yet, try testing your own programs. If montecarlo simulations
are not that common, perhaps you'll be the first to publish your
findings...

V

 

[Lars Schouw]

Victor
Thank you for the nice and useful answer. |-)

tomshardware tests ....hardware I want to see how a monte carlo
simulation (software) runs on the same CPU but with a differerent OS
and compiler setting - build for 64 bit

Lars

 

[Gianni Mariani]

Victor
Thank you for the nice and useful answer. |-)

tomshardware tests ....hardware I want to see how a monte carlo
simulation (software) runs on the same CPU but with a differerent OS
and compiler setting - build for 64 bit

I think Victor is saying you're better off trying a newsgroup that is
more related to your question. comp.arch, or some MS group on 64 bit
stuff would probably get a useful response. He is right.

As an aside, my experience is that there is a significant speed-up but
this is not because of the inherant "64 bitness" but because going from
32 bit X86 ISA to AMD64 means the compiler has some more registers to
play with and does allows the compiler to do a much better job with code
generation. There is a hit when you start storing 64 bit pointers as
they use twice the memory. My experience with mips 32 bit (o32 and n32)
vs the 64 bit (n64) instructions was that n32 was significantly faster
than n64 on some applications (usually memory bound apps).

In the case of AMD64 vs X86-32, I don't think I've ever seen a
performance regression going to 64 bit.

Every performance question ends up with this answer - "Test it". Only
you are qualified to do that.

 

[Lars Schouw]

Do you think that 64 bit performance is more OS related than compiler
related? So better to move to a more os specific group like W$indows,
Un$x or Lin**x?

I don't really have time to test it myself right now so it would be
great if someone already did it.

Lars

 

[Ian Collins]

As an aside, my experience is that there is a significant speed-up but
this is not because of the inherant "64 bitness" but because going from
32 bit X86 ISA to AMD64 means the compiler has some more registers to
play with and does allows the compiler to do a much better job with code
generation.

Agreed, I've seen the same.

 

[Gianni Mariani]

Do you think that 64 bit performance is more OS related than compiler
related? So better to move to a more os specific group like W$indows,
Un$x or Lin**x?

I don't really have time to test it myself right now so it would be
great if someone already did it.

If your code is computationally intensive (which is sounds like it is),
it's most likely bound by the algorithms and the compiler.

 

[Ian Collins]

Do you think that 64 bit performance is more OS related than compiler
related? So better to move to a more os specific group like W$indows,
Un$x or Lin**x?

Yes.

The topic has been discussed at length on Solaris forums, so I guess it
has on other OS specific groups.

 

[Lars Schouw]

Ian

I guess the most Solais C++ code is still running on UltraSparc
processors are a little bit different than Intel in arcitecture.
I will have a look in the Solaris usenet group anyway.

Cheers
Lars

 

[roberts.noah]

Agreed, I've seen the same.

I don't know that there is any noticable improvement using 64 bit code
over 32 bit code. In fact I might go as far as to say 32 bit code is
faster but that could just be my perception. 32 bit systems are
certainly more stable right now.

 

[Ian Collins]

Ian

I guess the most Solais C++ code is still running on UltraSparc
processors are a little bit different than Intel in arcitecture.
I will have a look in the Solaris usenet group anyway.

Not quite, Solaris has been running on AMD64 for well over a year now
and the native compilers (used to build the OS) are well tuned for 64
bit. There was a lot of interest in performance when it started shipping.

 

[Ian Collins]

I don't know that there is any noticable improvement using 64 bit code
over 32 bit code. In fact I might go as far as to say 32 bit code is
faster but that could just be my perception. 32 bit systems are
certainly more stable right now.

Have you run comparisons?

In Unix land, 64 bit OSs have been shipping for a while (years in some
cases) and are no different in stability to 32 bit ones.

 

[roberts.noah]

Have you run comparisons?

Nope. Just perception like I stated.

In Unix land, 64 bit OSs have been shipping for a while (years in some
cases) and are no different in stability to 32 bit ones.

OS is Linux. Using the 64 bit version introduces many hassles that
aren't problems in 32 bit version.