[molpro-user] CUDA Support

[Ismail, Mohd F.]

Thank you.  I have applied for the trial license and will take a look at it.

--Farid


________________________________________
From: Gerald Knizia [knizia at theochem.uni-stuttgart.de]
Sent: Thursday, January 26, 2012 1:45 PM
To: Ismail, Mohd F.
Cc: molpro-user at molpro.net
Subject: Re: [molpro-user] CUDA Support

On Thu, 2012-01-26 at 06:44 +0000, Ismail, Mohd F. wrote:
> Hi, I saw that molpro contains cuda-accelerated code.  Can someone
> list in more detail which part(s) of molpro is(are) accelerated by
> cuda?  Is there a benchmark of the speed improvement for molpro?

In the current release version (2010.1) there is some cuda DFT code, but
it is best described as "very experimental" and definitely not
recommended for use in practice. But the upcoming Molpro version, which
will be released within the next two months, will have GPU-accelerated
exact exchange treatment of DF-MP2 and (density-fitted)
Hartree-Fock/Kohn-Sham. These new codes actually work very well (but
neither DFT nor gradients can use the GPU yet).

Note, however, that GPU-acceleration in general is somewhat limited in
scope: Unless you have dedicated scientific "GPGPU" hardware, the units
are not actually much faster than CPUs for double precision
arithmetic[1]. But there will be plenty of improvements to the CPU-based
versions of DF-HF/DF-KS (including gradients), and DF-MP2, so there will
anyway be advantages in performance.


[1] Even in terms of raw FLOP performance, neglecting things like the
difficulty in programming them and host<->gpu data communication, etc.
E.g., a Geforce GTX 580 apparently has around 200 GFLOP/s (double
precision). My quad-core i7 notebook has 70.
--
Gerald Knizia