TC schrieb:
> Could you run the benchmark with 32-bit PoyRay binaries? It should not take 
> too much time to download - and less then two minutes for you to run. ;-)

Not really:

- The 32-bit beta won't install unless I uninstall the 64-bit version 
first (which, as you can guess, I'm rather reluctant to do ;-))

- Although I compile patched versions of beta.33 almost daily, and it 
wouldn't be a big deal to just run a "non-installed" 32-bit binary once 
I got it built, I never tried before on my 64-bit machine. So, for 
instance, I didn't have 32-bit boost libraries ready.

Well, you can see that I'm using past tense here, so obviously I went 
that extra mile ;-)

Note however that the results must be used with caution:

- Windows XP x64 is a native 64-bit operating system; I guess any 32-bit 
application will suffer some slowdown due to overhead for switching 
between 32-bit and 64-bit every time OS functions are called (such as 
mutexing and other multithreading related stuff, file access, etc). So 
part of the speed difference could indicate not an advantage of 64 bit, 
but a disadvantage of running a 32-bit app on a 64-bit OS.

- I found that the SSE version didn't compile in Visual Studio 2005 for 
some obscure optimization-related issue, so I had to turn off some 
optimization option. (The non-SSE version does compile "as is", but 
would of course distort the results because the 64-bit version always 
uses SSE.)

I also ran a "home-brewn" 64-bit build, to make sure I'm not comparing 
compilers (official releases are built with Visual Studio 2008) or the 
effect of some experimental patches presently in my working code.

The results I got are as follows:

32-bit:     150 s
32-bit SSE: 123 s
64-bit:      74 s

So at least it can be concluded that using the 64-bit binary /does/ give 
a significant advantage /when running a 64-bit OS anway/.

Given the impressive observed difference, I also /think/ that the 
overhead from running a 32-bit app on a 64-bit OS and/or lack of some 
optimization can't possibly account for all of it, and that a "64-64" 
system indeed gives an advantage over a "32-32" system. However, I can't 
guarantee.

Post a reply to this message