Hello,
I recall an email conversation with Ross Walker a long time ago - back in 2014. We discussed the installation of Amber12 for GPUs. Looking at my emails, I see that at the time Ross didn't think it was worth compiling the CUDA version with the Intel compilers. Given that the CUDA version is designed to spend as much time as possible on the GPU cards then that does make a lot of sense.
We're now working on Amber16/AmberTools17 on a new cluster (OS RHELS 7.4). Building CUDA enabled Amber with the Intel compilers (v17.0.4) works, but there are many test failures. Actually, I don't recall a single pass. Most of the comparisons aren't showing significant differences, but failures concern me. On the other hand my GNU build (v4.8.5) is perfect. I am, by the way, using CUDA v9.0. Is it still true that it isn't really worth building the CUDA enabled version with the Intel compilers? If this is the case then I'm happy to go with the GNU compiled version, but I thought it best to check.
Could someone advise on this one, please?
Best regards,
David
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Feb 21 2018 - 08:30:02 PST
