Hi Filip and Jan-Philip,
Yes the performance regression from 4.1 is fixed in 4.2. We will be having a
patch over the next month or so that adds Kepler support to AMBER 12 and
this will then require 4.2 as a minimum so you might want to consider
switching to 4.2 now.
All the best
Ross
> -----Original Message-----
> From: filip fratev [mailto:filipfratev.yahoo.com]
> Sent: Wednesday, May 16, 2012 4:54 AM
> To: AMBER Mailing List
> Subject: Re: [AMBER] Is the performance issue of CUDA 4.1 compilers
> resolved in CUDA 4.2 compilers?
>
> Hi,
> It seems that issue was resolved or at least the speed of Amber12-cuda
> with CUDA 4.0 and 4.2 is the same. I tested this last night.
>
>
> All the best,
> Filip
>
>
> ________________________________
> From: Jan-Philip Gehrcke <jgehrcke.googlemail.com>
> To: AMBER Mailing List <amber.ambermd.org>
> Sent: Wednesday, May 16, 2012 2:45 PM
> Subject: [AMBER] Is the performance issue of CUDA 4.1 compilers
> resolved in CUDA 4.2 compilers?
>
> Huhu,
>
> as Ross pointed out before, there is a performance regression in CUDA
> 4.1 compilers (cf. mail below). On the NVIDIA website, in the
> meanwhile,
> CUDA Toolkit & SDK are available in version 4.2.9:
> http://developer.nvidia.com/cuda-downloads
>
> Unfortunately, I cannot find a changelog. Does anyone know for sure if
> the discussed issue was resolved with this 4.2.9 release?
>
> Thanks,
>
> Jan-Philip
>
>
>
> > Hi Ashraf,
> >
> > You can use CUDA 4.1 but it is approximately 10 to 15% slower than if
> you
> > used 4.0 due to performance regression bugs within the 4.1 compiler.
> Our
> > advice at present is to use the CUDA 4.0 compilers until such a time
> as the
> > compiler bug is fixed. Hopefully this will be with the release of the
> 4.2
> > compiler.
> >
> > All the best
> > Ross
>
>
> _______________________________________________
> 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
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed May 16 2012 - 08:30:04 PDT
