Re: [AMBER] RTX 2080 slower then 1080 Ti with pmemd 18 ?

From: Ross Walker <ross.rosswalker.co.uk>
Date: Wed, 21 Nov 2018 15:43:35 -0500

Correction $AMBERHOME/src/pmemd/src/cuda/gputypes.h

> On Nov 21, 2018, at 15:37, Ross Walker <ross.rosswalker.co.uk> wrote:
>
> Hi Alain,
>
> You will need a patched version of AMBER to get full performance on RTX cards. That is still being worked on. In the meantime edit the following line in $AMBERHOME/src/pmemd/src/cuda/gpu_types.h
>
> static const int PMENONBONDFORCES_THREADS_PER_BLOCK = 640;
>
> Change to
>
> static const int PMENONBONDFORCES_THREADS_PER_BLOCK = 512;
>
> Then recompile. Note this will run slower on non RTX2080(TI) cards so you will need multiple executables if you have mixed hardware.
>
> Also use CUDA 9.2 for the RTX cards.
>
> All the best
> Ross
>
>> On Nov 21, 2018, at 14:24, Alain Chaumont <chaumont.unistra.fr <mailto:chaumont.unistra.fr>> wrote:
>>
>> Dear all,
>> we currently are testing our new RTX 2080 cards (MSI Geforce Seahawk RTX 2080). However it seems that they are about 2% slower then our GTX 1080Ti Cards (Similar model).
>> However from the benchmarks on the AMBER webpage I find that the RTX 2080 should be about 10 - 20 % faster then GTX 1080Ti. Is this correct? If so any ideas why this is not the case for us? Is there some “trick” to use during compilation or so? Or are they done with some patches of AMBER which are not yet public?
>>
>> Thanks a lot in advance
>>
>> Alain Chaumont
>>
>> P.S. We are currently using CUDA 9.0 . The size of the systems is about 130 000 atoms.
>>
>>
>>
>> Universite de Strasbourg
>> 4, rue B. Pascal
>> 67000 Strasbourg
>> France
>>
>>
>>
>> _______________________________________________
>> AMBER mailing list
>> AMBER.ambermd.org <mailto: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 Nov 21 2018 - 13:00:04 PST
Custom Search