MathGroup Archive 2009

[Date Index] [Thread Index] [Author Index]

Search the Archive

Re: Performance on 64 bit OS

  • To: mathgroup at smc.vnet.net
  • Subject: [mg103104] Re: [mg103075] Performance on 64 bit OS
  • From: John Fultz <jfultz at wolfram.com>
  • Date: Tue, 8 Sep 2009 05:57:32 -0400 (EDT)
  • Reply-to: jfultz at wolfram.com

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.

Sincerely,
 
John Fultz
jfultz at wolfram.com
User Interface Group
Wolfram Research, Inc.



On Mon, 7 Sep 2009 02:35:04 -0400 (EDT), fd wrote:
> 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
> windows)
>
> 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)
>
> Best





  • Prev by Date: Re: Optimizing a function with a list as a parameter
  • Next by Date: Re: How does Eigensystem works ?
  • Previous by thread: Performance on 64 bit OS
  • Next by thread: Re: How reliable are mathematica's special functions? (Hypergeometric