Re: which Intel mobile dual core processor is best for floating point scientific comp

On Aug 18, 7:09 pm, Luna Moon <lunamoonm...@gmail.com> wrote:
> On Aug 18, 4:52 pm, Alfredo Buttari <alfredo.butt...@gmail.com> wrote:
>
>
> 
> 
> 
> 
> 
> 
> 
> 
>
> There are two arguments about using a single core in a dual-core
> system.
>
> You and many people say the program running on a single core in a dual-
> core system will be slower than a standalone single core.
>
> But I've heard people say they are different numbering system now, and
> although the GHZ number is lower, actually the performance is higher,
> for a single core in a dual-core system vs. a standalone single core.
>
> -----------
>
> Moreover.
>
> My primary concern is speed for scientific computation.
>
> For that people say desktop is better choice. They suggest me to lower
> my expectation(hence price) on laptop and buy desktop for computing.
>
> So here are the budget planning:
>
> 1. A very lousy laptop + upgrade my existing desktop(P4 3.2GHz), which
> is actually repurchasing a new desktop;
>
> 2. Focus on a high end laptop + live with my current P4 3.2GHz
> desktop.
>
> Which approach is the best in terms of cost and performance overall?
>
> Thanks

Luna,
well, the thing as actually much more complex than that. Scientific
computing may be roughly split in two categories performancewise:

1) memory bound operations: this includes all the sparse computations
and more
2) cpu bound: pretty much all the dense computations

obviously all the software in 1) will benefit from a faster bus and,
yes, modern multicore processor usually have a faster bus (even though
the two things are not quite related). For 2), since the floating
point units of a dual core (say, a core 2 duo) are pretty much the
same of those in a Pentium IV, a lower clock frequency means lower
performance unless you can use multiple cores in parallel.
This said, it's ok to do the development on a laptop (and in fact I do
development on mine) but not to do the actual work. There so many
reasons why a laptop is not good for production...
The suggestion to buy a lower end laptop for the development and a
desktop for doing the true runs perfectly makes sense to me.
Also, note that Mathworks has a parallel Matlab version. You may want
to give it a try even though it requires some parallel programming
knowledge.

Alfredo

Alfredo Buttari <alfredo.buttari@gmail.com> wrote

> Scientific
> computing may be roughly split in two categories
performancewise:
>
> 1) memory bound operations: this includes all the sparse
computations
> and more
> 2) cpu bound: pretty much all the dense computations

It's true that scientific computing splits into these two
categories, but sparse x=A\b is CPU bound, not memory bound
(except for sparse QR).

There is a BLAS performance bug that limits the sparse x=A\b
multithreaded performance (see Cleve's Corner), but this
will  be fixed when the BLAS are fixed.  When that occurs,
sparse x=A\b will benefit from multithreading.

For more details, see:

http://www.cise.ufl.edu/research/sparse/cholmod/
http://www.cise.ufl.edu/research/sp...od/blasbug.html
[url]http://www.mathworks.com/company/newsletters/news_notes/june07/clevescorner.html[/
url]

> since the floating
> point units of a dual core (say, a core 2 duo) are pretty
much the
> same of those in a Pentium IV, a lower clock frequency
means lower
> performance unless you can use multiple cores in parallel.

I have a Core 2 Duo E6700 overclocked to 3.5 GHz and I can
tell you that there is a huge speed increase over running a
pentium 4 at a similar clock frequency. I have used the
bench function in Matlab as a way to compare computing
power, and I am getting an overall score of 70 with a low
end graphics card (The 3d graphics part is included in the
score). The other parts of the bench (LU FFT ODE SPARSE) are
much more important to me, and my times beat the pentium 4
3.0 GHz times by a factor of between 2 and 4 across the
board. There is obviously a lot more work being done per
clock cycle with these new Core 2 Duo processors. For
running Simulink models, my results are typically 4 - 5
times faster than the pentium 4 machines that my co-workers
are using.

Reed