Gpu Computing

Previous: Leeenux V3

Next: Back Again

created: 1278946678|%e %B %Y, %H:%M
TAGS: geeky
Back to blogging, before the summer break. A friend of mine send me an interesting mail about GPU computing. Yes, GPU (as for Graphical Processing Unit) and not CPU. The mail is below:

On Thursday I went to London to learn about GPU (graphics processing
unit) accelerated calculations. GPUs seem to be the future in high
performance computing they can provide computing power (per gigaflop)
10 times cheaper and using 20 times less energy than conventional CPUs
(=> less electricity and air conditioning running costs). This is
achieved through a highly parallel architecture; each graphics card
can have up to 512 cores and one can put up to 4 GPUs in a single
desktop. Connecting 2000 CPUs together and achieving good scaling
would require huge investments in special high speed interconnects
(e.g. infiniband) that are far too expensive for most universities and
thus only used in supercomputing centres. This is why on Lucky (where
the nodes are connected with gigabit Ethernet) you can only hope for
good scaling using the cores in a single node, so max 8 cores. Using a
desktop with 4 GPUs, you have 2000 GPU cores in one machine using
shared memory, so you can run parallel jobs on 2000 cores without the
need for expensive high speed interconnects.

While not all applications have been rewritten to work on GPUs, there
is a lot of work going on and progress being made. See for details.

MD codes which now support GPU acceleration include:

NAMD 2.7 Beta 2 including CUDA Acceleration
HOOMD: Highly Optimized Object Oriented Molecular Dynamics
GROMACS using OpenMM
ACEMD Bio-molecular Dynamics Package

For quantum chemistry there is a Terachem, written especially for GPUs
which offers up to 1000 x speed up using a GPU compared to using a
single CPU using the GAMESS QC package. Unfortunately, the
functionality is limited at the moment; it can only handle S and P
orbitals and can only do basic single point calculations,
optimisations, transition state searches and molecular dynamics. No TD
or CI for the moment see

However, GAMESS (US) is currently working on a GPU implementation
which should be able to do CI and TD calculations, but it may be some
time before it is ready.

For those of you who use Matlab, there is also a GPU accelerated
version of Matlab see to get
an idea of the speedups which can be achieved.

If anyone wants to buy a GPU personal supercomputer, you can find them
at Once you put a
reasonable amount of RAM on the machine and add a 2050 GPU, it's about
5000 euros for one machine. But remember this can do an AMBER MD
simulation at the same speed as 64 cores on a supercomputer, which
would cost ~ 50000 euros. And you don't need a special air-conditioned
room, you can put it under your desk.

Not bad eh?

Bookmark and Share

Rate this post:

rating: 0+x

Comments: 0

Unless otherwise stated, the content of this page is licensed under Creative Commons Attribution-ShareAlike 3.0 License