VMD-L Mailing List

From: Paweł Kędzierski (pawel.kedzierski_at_pwr.wroc.pl)
Date: Tue Sep 27 2011 - 09:29:31 CDT

Dear VMD-users,
I am cc'ing this answer to the list in case someone might find it useful
(or would like to add/correct something).

W dniu 27.09.2011 01:05, Dean Thorson pisze:
>
> Let me preface this email by saying that I am the father of a college
> student that uses NAMD and VMD. I found your exchange on VMD-L as I
> was doing research for my daughter. I got a call yesterday from her
> saying that her current laptop was failing quickly, that she needed a
> new one and could I help find one that would be appropriate for normal
> college activity and be better suited to run NAMD and VMD when she is
> not using the super computer center. I was hoping that the University
> of Illinois site would have recommendations but I couldn’t find
> anything current. Talking to people at computer stores has proven of
> no value. What I tried to find was real experience with current
> laptop hardware configurations and NAMD / VMD. Your VMD-L question
> was one of the few I found.
>
Dear Dean,
Of course everything depends on the money you wish to spend, on the OS
you use, and on your priorities. I will try to be general, anyway :-)

As for VMD and NAMD, the first one wants good OpenGL support, while the
second would benefit most from multi-core CPU; but there are CUDA
versions of NAMD which could use NVidia graphics to speed up calculations.

VMD already offers significant support for (some) calculations using
CUDA on GPU, and support in NAMD appeared recently (I haven't tested it
myself yet). This kind of calculations is currently only possible with
the proprietary CUDA libraries from NVidia, and this is the very reason
that NVidia cards are recommended for VMD and NAMD. With a decent card,
such calculations may be a few to several tens times quicker than on
central processor, but these numbers were obtained using desktop setups
or even high-end power-hungry workstations.

For the info below concerning NVidia please accept my disclaimer as it
is what I remember from various sources, from the time I did my research
on this subject a few months ago.
NVidia offers two lines of GPUs available for laptops: the
"professional" Quadro cards, and the rest is for mass-market of gamers.
Actually they seem to be powered by the similar chip architectures but
sometimes stripped down in hardware (like having some processor cores
disabled) or tied to specific drivers. The Quadro cards are expensive
and their performance is tuned against business and scientific
applications, while they perform poorly (or even don't work) with games.
They are also relatively infrequent and may limit your choice to
relatively few top-of-the-shelf machines.
The gaming cards might have more memory, more cores and higher clock for
lower price, but they performance with CUDA is reduced because only some
of the GPU cores are available for CUDA calculations. I have also read
somewhere, that some specific OpenGL functions (useful for pro apps and
not for gaming) are unusually slow with cards other than Quadro or Tesla
branded ones. I do not know if or how it may affect VMD or NAMD.
Taking the above into account, I looked for a laptop with NVidia but not
necessarily something high-end, to be able at least to setup and test
GPU calculations prior to running them on something better.

Personal experience:
I am very pleased with my i7 SandyBridge CPU as I already tested
calculations using 8 cores in parallel (4 physical with HT) and even
with reported 8x100% CPU usage the machine was still very responsive for
interactive work, I could even watch fullscreen video without any
glitches (note this was on Linux).
The weakest point on the Dell XPS2 seem to be the cooling system,
because I during such calculations I observed regular (short) drops from
100% to 0% CPU usage for a half or all of the cores for no apparent
reason. These accounted only for some small fraction of time so it
wasn't that bad. These could be IO related bottlenecks, but I think it
is related to inadequate cooling since with just two cores on 100%CPU
the fan was already working on max rpm. It might get worse in case both
the Nvidia chip and the CPU are on full steam (I have not tested it like
this), but I believe that even in this case a quad-core should do
parallel calculations faster than a two-core processor.

 From my own experience with VMD on this new laptop, the Intel
SandyBridge i7 integrated graphics is 30-50 times slower with OpenGL
display than NVidia GT540M (measured with VMD FPS display). I would say
that for visualization the integrated graphics is not recommendable, as
even small proteins with high detail settings bog down the performance
to 1-5 FPS and the movement gets very jerky. That said, it is at least
usable with small systems and not too demanding graphical representations.

My previous rather old laptop had an AMD Radeon 9000M card with 32MB of
RAM and it worked quite well with VMD for display with modestly sized
structures and with default detail settings (up to a few tens of
thousands of atoms). It was however unable to render surfaces from
volumetric data like electron density or electrostatic potential, due to
insufficient graphics memory. Right now it seems that AMD chips offer
more performance per dollar than NVidia ones when comes to 3D graphics,
so if your daughter does not need any CUDA support, an AMD card could
work well.

Another word of caution:
Currently it is hard to find a laptop with i3-i7 CPU and NVidia GPU
without so-called Optimus technology. As you said nothing about the OS,
I assume Windows 7, and to my knowledge this is the only system where
programs can use the NVidia GPU with Optimus out of the box.

Since I am mostly Linux user, I hesitated with the purchase until I
found a successful solution for Linux:
http://www.martin-juhl.dk/2011/05/optimus-on-linux-problem-solved/
The offspring of this idea are two independent projects (bumblebee and
ironhide) but neither one is mature yet so this is something for
advanced users.

> If you are willing, I would be interested in how you selected your
> hardware configuration and any recommendations on what to consider.
>

I opted for quad-core Intel CPU and modestly priced (that is, gaming
level) NVidia card with many CUDA cores and large memory. I do not know
how large calculations your daughter will run but a lot of RAM (6GB or
more) and HDD space is probably recommendable, too. Having a lot of
money to spend you may want to get an SDD drive for the sake of speed
and robustness on the go, but in this case I would say that an external
large HDD with fast connection (USB 3.0 or at least eSATA) is essential.
For selection of a specific GPU, you may find this page useful (the
description is for GT540M but on the left you will find a long list of
links to other cards, sorted by performance):
http://www.notebookcheck.net/NVIDIA-GeForce-GT-540M.41715.0.html

Note that the better cards are only mounted in larger laptops due to
power consumption. My DELL XPS is declared a 15 inch machine (15,6''
actually) but it is rather comparable to 16'' ones and seems relatively
HUGE in comparison with my old 14'' one having good old golden-aspect
ratio display. It weighs about 3kg itself but its 120W power supply is
kind of a brick (probably another kg or more) and this together is quite
a lot to carry around the campus, especially for a girl.

And running calculations on battery is not recommended: with all CPU
cores running on 100% the battery life is 30-40min, compared to about
4h without load.

> Thank you for any information you can provide.
>
You are welcome.
Greetings,
Pawel Kedzierski