On Sun, Jul 16, 2017, James Starlight wrote:
>
> I have found a lot of information regarding customization of the
> desktops with 2-3 GPUs set up for the AMBER. However I have fond
> nothing for the GPU-based server solutions, e.g when I have a
> "central" cpu node communicated with the several GPUs nodes via
> infiniband. Could someone post some examples of such servers with a
> good benchmarks for MD simulations and its technical details? Should
> the performance of such server (measured per one GPU node) be better
> than in case of the regular desktop, assuming that for the test I
> submit one simulation for one GPU?
Be sure to read (if you have not already) this page:
http://ambermd.org/gpus/recommended_hardware.htm#hardware
There is a "build you own machine" link as well as a lot of descriptions of
setups for GPU-based servers, with 8 or more GPU cards. Infiniband is not
relevant here: the GPU cards will communicate directly with the CPU via a PCI
bus.
All the MD work is done on the GPU, so the performance per GPU will depend on
the GPU card you choose (see the benchmarks page), and will not depend on
whether you have a "desktop" or a "server".
....dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Sun Jul 16 2017 - 11:00:03 PDT
