Re: [AMBER] Different convergence in pmemd.cuda against CPU

From: Hiromasa WATANABE <hi-watanabe.hpc.co.jp>
Date: Mon, 22 Dec 2014 15:18:02 +0900

Dear Dr. Ross,

I'm sorry to be late for reply. Yes, your summary is correct.

As your advice, last week I tried with Amber14, and it's OK.
Here's my summary:

AMBER12 (CUDA pmemd)
x1 0.9839g/cc
x2 0.9837g/cc
x4 0.9530g/cc
x8 0.9534g/cc

AMBER12 (CPU)
sander x8 0.9838g/cc
pmemd x8 0.9836g/cc

AMBER14 (CUDA pmemd)
x1 0.9837g/cc
x2 0.9837g/cc
x4 0.9837g/cc
x8 0.9837g/cc

Thanks for your advice!

Best regards,
WATANABE


> Hi Hiromasa
>
> Ahhh - so 4 GPU and 8 GPU runs converge to the wrong density while 1 and 2
> GPU runs work correctly. Is that the correct summary? - That indeed does
> suggest a bug and probably slipped through since 4 and 8 GPU runs are
> rarely tested since it very rarely (with AMBER 12) gives any speedup. Do
> you have a copy of AMBER 14 - could you possibly try running with that and
> see if you see the same behavior please - this will help in tracking down
> the problem.
>
> Thanks,
>
> All the best
> Ross
>
>
> On 12/15/14, 8:42 PM, "Hiromasa WATANABE" <hi-watanabe.hpc.co.jp> wrote:
>
>> Hi,
>>
>> We have run TIP3P water system (300K, 1bar) using AMBER12(patch 21).
>>
>> pmemd.cuda(x4 or x8 parallel) converge density to 0.95g/cc,
>> whereas sander(x8), pmemd(x8) and pmemd.cuda(x1, x2) converge to 0.98g/cc.
>>
>> Please suggest any ideas to us.
>>
>> Best regards,
>> WATANABE
>>
>> --
>> Hiromasa WATANABE
>> Manager, Ph.D.
>> HPC Dept., Technology Gr., HPC SYSTEMS Inc.
>> Head office: LOOP-X 8F, 3-9-15 Wangan, Minato-ku, Tokyo, Japan 108-0022.
>> Email: hi-watanabe.hpc.co.jp
>> www.hpc.co.jp
>>
>> _______________________________________________
>> AMBER mailing list
>> AMBER.ambermd.org
>> http://lists.ambermd.org/mailman/listinfo/amber
>
>
>

-- 
Hiromasa WATANABE
Manager, Ph.D.
HPC Dept., Technology Gr., HPC SYSTEMS Inc.
Head office: LOOP-X 8F, 3-9-15 Wangan, Minato-ku, Tokyo, Japan 108-0022.
Email: hi-watanabe.hpc.co.jp
www.hpc.co.jp
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Sun Dec 21 2014 - 22:30:03 PST
Custom Search