[SciPy-user] 64 bit and 32 bit on Ubuntu
Sun Oct 12 16:04:34 CDT 2008
On Sun, Oct 12, 2008 at 7:14 PM, Nathan Bell <firstname.lastname@example.org> wrote:
> On Sun, Oct 12, 2008 at 2:45 PM, Kurt Smith <email@example.com> wrote:
>> for numerical work, specifically under scipy, etc. I'm not well
>> versed in the low-level pros and cons of 64 bit vs 32 bit, esp. as to
>> how it affects scipy performance. Can you enlighten me? I"m on an
>> Intel Core 2 Duo, 3.16 GHz, E8500, 4 GB RAM. Will using 64 bit yield
>> great benefits running scipy/numpy code automatically, or will it
>> require fine-tuning on my part? Will 64 bit yield benefits besides
>> having a larger address space (besides the fact that my 4 GB RAM is
>> still addressable by 32 bits, anyway)?
> I wouldn't expect any significant performance difference. The x86-64
> architecture does bring a few improvements such as more registers and
> (guaranteed?) SSE2 support, but the benefit of these depends on the
> nature of the application.
> Something in the neighborhood of 20% faster for general purpose codes
> should be typical. OTOH I've found that some pure Python code that
> makes heavy use of dictionaries runs slower, possibly owing to the
> fact that 64-bit pointers reduce the effective CPU cache size.
Is this 20% from your own experience, or are there timings out there?
I've looked around, but the comparisons are pretty old and made when
very few 64 bit apps were around, so 32 had the day. Anyone aware of
comparisons between 64 and 32 bit performance for numerical codes?
> Also, you wouldn't be able to address *all* 4GB RAM in a 32-bit OS.
> Both Windows and Linux reserve some part of the upper addresses for
> the OS, so a process will be limited to something less (3GB to 3.5GB
> Keep in mind that 64-bit Linux is approximately 2x cooler than 32-bit
> Linux, so the choice is easy IMO.
This temperature difference is really surprising to me, and I'm not
able to find anything that talks about it in detail. Anything more
you can say as to why this would be the case?
More information about the SciPy-user