Re: [AMBER] CUDA and driver version for Nvidia Titan Black and Titan Z GPUs cluster

From: Ryan Novosielski <>
Date: Mon, 10 Sep 2018 16:33:09 -0400

Hash: SHA1

On 09/10/2018 11:20 AM, David A Case wrote:
> On Fri, Sep 07, 2018, Karolina MitusiƄska (Markowska) wrote:
>> I would like to ask you about cluster configuration. We are using
>> Amber14 on our Nvidia Titan Black and Titan Z clusters. We are
>> now running on CUDA 6.5 and 343.19 driver. Because we are also
>> running on Ubuntu 14.04 and it's a bit old fashioned now, we want
>> to upgrade to Ubuntu 18.04 and also (for other reasons) to
>> Amber18.
>> Can you suggest the driver and CUDA version that will suit us?
>> And will that configuration run smoothly with Amber14 also?
> It may (or may not) be involved to have a single configuration that
> works well with both Amber14 and Amber18. The GPU code in Amber18
> is *much* better than that in Amber14; are you sure you need the
> latter?
> Amber18 will work fine with Ubunutu 18.04, using gcc7 and CUDA
> version 9.2, and current NVIDIA drivers (not sure what the number
> is).
> To support Amber14 (roughly): use the "alternatives" apt-get
> capability to install whatever compilers you currently have on
> Ubuntu 14.04. Then, you can try to use those compilers to compile
> Amber14's pmemd.cuda, and see if that executable will work with the
> current CUDA and NVIDIA drivers. If it works, (and, as far as I
> can tell, there is a good chance it will) you are set to go.
> (You *could* try compiling Amber14 pmemd.cuda with gcc7/gfortran7,
> perhaps fixing things that are broken.)
> If not, I don't know how to proceed without having a dual-boot or
> pair of virtual machines, one with your current configuration and
> one with Ubuntu18.04. Maybe someone on the list has a better
> idea.

This is the sort of situation we handle via modules; currently, I
think most people would recommend Lmod, but there's the Tcl based
environment package as well. You can have one copy compiled with one
and one with the other. I think the real thing to look out for is that
you have a driver version that will work with both. So far, I've not
had a problem with backward compatibility and NVIDIA drivers (we use
M2070 cards all the way up to P100 cards with the same driver -- I've
not tried very old CUDA with it though).

- --
 || \\UTGERS, |----------------------*O*------------------------
 ||_// the State | Ryan Novosielski -
 || \\ University | Sr. Technologist - 973/972.0922 ~*~ RBHS Campus
 || \\ of NJ | Office of Advanced Res. Comp. - MSB C630, Newark


AMBER mailing list
Received on Mon Sep 10 2018 - 14:00:01 PDT
Custom Search