Re: [AMBER] suggestion for a GTX 1080

From: Ross Walker <ross.rosswalker.co.uk>
Date: Mon, 13 Nov 2017 13:40:03 +0100

Hi Ilhan,

You are misunderstanding what I mean here. Any GTX1080 should work fine and should give bitwise identical results to any other GTX1080. There are no concerns about accuracy between different GPU cards. PMEMD.cuda's precision model was designed from the ground up to work identically on Geforce, Quadro and Tesla GPUs. The issue I am referring to is that sometimes, although it tends to be very rare (<1%) these days in my experience, Geforce cards can be faulty and give incorrect results. This often manifests in simulations being unstable and crashing regularly. The problem is most likely to occur with overclocked cards, particularly those that aggressively overclock the memory. I've also seen it regularly with second hand cards (ebay, Amazon market place etc) but I suspect this is people trying to offload cards they know to be unreliable.

Ultimately if you avoid the very highest overclocked cards in a series you should be fine. It's easy to test the card once you receive it. Just run a large AMBER simulation multiple times on the GPU with identical input. The output in each case should be completely identical.

The other thing it is often good to avoid are the cards with custom cooling - lots of brightly colored fans etc. These coolers are fine if you only plan to have one card in the machine but if you plan to have multiple cards the top down nature of the airflow can mean that the cards do not get effectively cooled when another card is beside them. Again this is a bigger issue with heavily overclocked cards since they will tend to run hotter and so require more aggressive cooling.

I hope that helps.

All the best
Ross

> On Nov 13, 2017, at 09:29, ilhan.yavuz <ilhan.yavuz.marmara.edu.tr> wrote:
>
> Hi, thanks for the suggestions. It's good to know that any GTX1080
> will work. Although, not the performance but the accuracy of the
> produced data
> might be the problem. It shouldnt be a problem, though, if only
> four/five
> digits after the decimal point matter for an MD.
>
> Thanks again, Ilhan
>
> 2017-11-13 04:34, Scott Le Grand yazmış:
>> I managed to buy a 1080TI off the NVIDIA web site a few weeks back.
>>
>> On Sun, Nov 12, 2017 at 4:06 PM, Ross Walker <ross.rosswalker.co.uk>
>> wrote:
>>
>>> Hi Ilhan,
>>>
>>> In principle any GTX1080 will work. Some just tend to be more reliable
>>> than others. Also the custom fan design ones tend to only be good for
>>> single GPU configurations but if that is what you plan then most
>>> models
>>> should work fine. I tend to recommend avoiding the super overclocked
>>> versions but the 1080 silicon is now sufficiently mature that those
>>> are
>>> probably okay as well these days. Just be sure to test it upon
>>> arrival.
>>>
>>> This one will probably work well for you and is fairly cheap:
>>> http://amzn.to/2AG6HPu <http://amzn.to/2AG6HPu>
>>>
>>> All the best
>>> Ross
>>>
>>>> On Nov 11, 2017, at 19:41, ilhan.yavuz <ilhan.yavuz.marmara.edu.tr>
>>> wrote:
>>>>
>>>> Dear all, can anyone suggest me a graphics card with GTX 1080
>>>> to use for amber, other than the one given in AMBER's Supported GPUs
>>>> page? It
>>>> would be better if it has the same quality and a similar price.
>>>>
>>>> Regards, Ilhan
>>>>
>>>> _______________________________________________
>>>> 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
>
> _______________________________________________
> 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 Mon Nov 13 2017 - 05:00:05 PST
Custom Search