Performance on 64 bit OS

Skip to first unread message


Sep 7, 2009, 2:34:54 AM9/7/09
Hi All

I recently run a BenchmarkReport[] on my machine at work (Core 2 duo
@2.66GHz, 4GB DDR2 RAM on windows xp 32 bit, NVIDIA graphs with
512MB). I wanted to make a comparison with my laptop, a much newer
macbook pro (Core 2 @2.53, 4GB DDR3 RAM, NVIDIA 256 on Snow leopard).

Though my desktop machine is slightly faster, it was running a 32 bit
operating system, so I was expecting the mac to be at least marginally
better. My surprise was when I got an overall of 2.96 with windows XP
against 2.41 for the Mac (after 4 tests, mean time 35.73 mac x 29.16

Individually the biggest difference appears to be on test 5 -
Elementary functions - (2.52 Windows x 7.38 on a Mac). The mac was
better at tests 8,7 and 6 (Matrix arithmetic, Large Integer
multiplication and gamma function) with scores of respectively 3.69 x
2.86, 0.94 x 0.53 and 0.53 x 0.29, which makes sense.

Is this expected? or am I doing something wrong on my 64 bit operating
system machine?

Another question is whether Mathematica will take advantage of OpenCL
technology (or any GPGPU technology as windows 7 is promising)


John Fultz

Sep 8, 2009, 5:56:22 AM9/8/09
It's a myth that 64-bit CPUs and operating systems are or ought to be faster
than an equivalent 32-bit CPU or operating system. In reality, there are some
things that they can do faster due to the much larger internal registers and
other mechanisms provided by the extra width. But on the down side, every
pointer is now 64-bits wide, which means that a lot of the data sitting in
memory just got bloated by a fair amount, causing more cache misses which
affects performance the other way.

So, if you're moving to 64-bit hardware/OS because you expect a performance
increase over 32-bit hardware, you might be making a mistake unless benchmarks
which are relevant to what you'll actually be doing prove you right. The one
thing you will most definitely get out of a 64-bit upgrade, though, is the
ability to address much more memory. This advantage, by itself, is enough for
many power users to switch.

I say all of this without commenting on your specific benchmark numbers. This
computation is out of my area, and I can't say whether the numbers you'reseeing
are reasonable. You should interpret my remarks above as general statements
rather than an attempt to specifically explain the benchmarks you've observed.

As for your second question, Mathematica does not presently take advantage of
OpenCL or any other GPU-based computation (aside from those used for
hardware-accelerated graphics rendering, of course). We're certainly
investigating these technologies, and that's all I can say for now.


John Fultz
User Interface Group
Wolfram Research, Inc.

Reply all
Reply to author
0 new messages