What kind of hardware works best for the Amber Molecular Dynamics Package?



We’re looking into buying hardware optimized for running amber, what hardware or specifications should we be paying attention to?
CURATOR: jpessin1


For local setups Amber’s hardware pages http://ambermd.org/intel/index.htm &
http://ambermd.org/gpus/index.htm#supported_gpus are a good start.


Other considerations: in general it is recommended to run Amber on GPU-accelerated nodes, for best performance. Should one have access to CPU-only nodes, suggestions are not to oversubscribe nodes; running one MPI task per core gives maximum performance (i.e., set ntasks-per-core=1). Curator: jpessin1


The Amber GPU pages have been updated, http://ambermd.org/GPUSupport.php. Specific GPU hardware is at: http://ambermd.org/GPUHardware.php. Basically most nVIDIA GPUs are excellent, however if you go to relatively old GPUs, newer versions of Amber do not work. Definitely one would want the latest Amber, Amber18 for both CPU and GPU performance updates. Amber is one of the fastest MD codes on GPUs but does not scale terribly well across multiple GPUs unless you have NVLINK and definitely not at all across multiple nodes, on both the CPU and GPU. However, there are a variety of ensemble methods, such as replica exchange MD, that can very efficiently take advantage of multiple GPUs or CPUs. Finally, depending on your nVIDIA EULA adherence tolerance level, the commodity GPU cards-- such as the 1080TI-- perform really well. For more information, e-mail the Amber mailing list or peruse the archive on ambermd.org. Note also that some of the XSEDE sights have GPU nodes, including bridges, comet, x-stream, and others.