Re: [AMBER] cuda version 5 using gnu V 4.9.1 issues

From: Ross Walker <ross.rosswalker.co.uk>
Date: Fri, 16 Jan 2015 12:16:43 -0800

Hi Fatima,

Did you download the correct version for your operating system? Redhat 7 and Centos 7 are GCC 4.8.2-16 and I've had CUDA 5.0 working fine with that. Note, you don't have to use CUDA 5.0 - v6.5 is fine if that's easier - you just lose about 5 to 7% or so in performance but it's probably not worth the hassle if it is a pain to use the older version of CUDA.

Note, I'd be very wary of bleeding edge compiler versions like GCC 4.9.1 - we've seen issues with it incorrectly compiling code (ends up giving the wrong answer) on OSX. This is why the Exxact machines are still using 4.4.7 - it's bullet proof. The 4.9 issue may just be restricted to OSX - these are the only machines I have that have the latest compilers (I tend to stick to RHEL or CENTOS for stability). I'd recommend running the test cases - particularly the serial and parallel CPU test cases and make sure they all come out good before trying GPU here. The issue on OSX with GCC 4.9 is with the CPU PMEMD TI code.

All the best
Ross

> On Jan 16, 2015, at 7:25 AM, CHAMI F. <fatima.chami.durham.ac.uk> wrote:
>
> Hi all,
>
> When I recompiled amber14 with cuda enabled I encountered the problem that cuda 5 requires an old version of the gcc (lower than 4.7) whereas I use 4.9.1 to stay up to date. Iíll start a rebuild with a lower version.
>
> I also notice that exxact machine spec the gnu GNU v4.4.7-3 was used with CUDA 5 ..
>
> thanks
> Fatima
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber


_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Jan 16 2015 - 12:30:04 PST
Custom Search