[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Fwd: Re: Opinion question (Core2 Duo)

On Wed, 2007-09-19 at 00:08 -0700, Francesco Pietra wrote:
> I have just completed unrestricted MP2/6-31G* energy procedure (after
> DFT/M05-2X) for a 98-atoms (first row) molecule in 19 hours with four-node
> amd64 Debian amd64 etch, NWChem suite. I understand (if I understand correctly)
> from your email that should I have had Core 2 I would not have had the time to
> take a coffee in between launching the MP2 procedure and getting the
> computational results. Interesting.

Xeon/Core2 cheats slightly - it has combined units for both adding and
multiplying on the chip - meaning if you do a multiplication immediately
followed by an addition, then it'll happen in 1 cycle instead of 2,
hence the "twice as many FLOPs" thing. Honestly, in desktop
applications, it doesn't matter much - but in matrix operations, as used
by most chemistry packages such as Gaussian or NWChem, it makes a full

So yes, you'd probably see a ~90% speed boost moving from an AMD64 to a
Core2 of identical clock speed.

If you have access to the NWChem source (you might not, I don't think we
have any login credentials lying around to check with) you would see
even bigger improvements with a commercial compiler and math library -
you may find a cheaper option to improve performance than buying a new
Core2 rig is to buy Portland C (or Pathscale C), and link against the
free AMD Core Math Library instead of conventional open-source
BLAS/LAPACK routines.

/ Jo Shields <jms@osc.ox.ac.uk> \
| Systems Manager,              |
\ Oxford Supercomputing Centre  /
       \   ,__,
        \  (oo)___
           (__)    )\
              ||--|| *

Reply to: