[gmx-users] Results of villin headpiece with AMD 8 core

[paul buscemi]

Dear M,

Yup the timestep is enormous, but the Gromacs demo used 0.005 ns.  At that point the 2700 and the Ti started flinging waters molecules off into the ozone. For the very short runs at 0.005 ~1400 ns/day was being hit.  Usually the ts I use  is 0.002 

Getting the ion and SOL concentration correct in the top is trickier ( for me ) than it should have been,   If you happen to  reuse  both  solvate and genion during  the build  keeping track of the top is like using a digital rubics cube..!  The charge  the villin was +1 because after I downloaded it from the pdb I removed all other water and ions - it just made pdb2gmx easier to work with.

 Interesting  that while adjusting ntmpi and ntomp made some difference the biggest influence came from rvdw, r coulomb cutoffs ,and rvwd-switch. They are apparently coupled in some manner. I was went as far as rvdw of 1.6 with much worse results. But not much is said about these parameters as compared to setting mpi and threads in regard to efficiency.  But this is just my inexperience showing

The 1080 scaled nicely with the 1080 ti,  these are really nice pieces of hardware. and you are correct, given the choice of increased processors vs  faster processors - choose the latter. I have the AMD OC to 4.0 GH and it runs the same model almost as fast as as 32 core AMD at 3.7 GHz.

I've run 300k DPPC models ( ~300 DPPC molecules ) and they run at ~15 ns/day in NPT.  And yes,  if you can send the pdb, top, and itps I’t would be interesting to compare the two AMDs. 

Best
Paul






> On Jan 11, 2019, at 3:27 PM, Wahab Mirco <Mirco.Wahab at chemie.tu-freiberg.de> wrote:
> 
> On 11.01.2019 19:55, pbuscemi wrote:
>> For those of you considering a workstation build and wonder about AMD processors I have the following results using the included npt and log intro for the villin headpiece in ~ 8000 atoms spc/e. The npt was run from a similar nvt ( 100000 steps ) . The best results were achieved with the simplest command line - letting Gromacs choose threads.
>> The system became unstable at dt =0.005 ns step. Note the close correspondence between rcoulomb, rvdw and cutoffswitch. Results compare favorably with the E5-2690+GTX Titan demo
>> http://on-demand.gputechconf.com/gtc/2013/webinar/gromacs-kepler-gpus-gtc-express-webinar.pdf (https://link.getmailspring.com/link/1547231722.local-ad2d5ea3-b061-v1.5.2-31660462@getmailspring.com/0?redirect=http%3A%2F%2Fon-demand.gputechconf.com%2Fgtc%2F2013%2Fwebinar%2Fgromacs-kepler-gpus-gtc-express-webinar.pdf&recipient=Z214LXVzZXJzQGdyb21hY3Mub3Jn)
>> Core t (s) Wall t (s) (%)
>> Time: 112.643 14.080 800.0
>> (ns/day) (hour/ns)
>> Performance: 1288.622 0.019
> 
> Hi Paul,
> 
> I couldn't avoid to test this on my R2700X box which has a
> GTX 1080 because this allows me to see the difference
> to the GTX-1080 Ti. According to your md.log, the boxes
> and the OS are very similar, I only had some problems first
> to get the old villin benchmark to run with your mdp-file
> (I added two chloride ions and changed the solvent to SPC/E
> to the original configuration). BTW. you used a rather
> large timestep in your test.
> 
> So, this would be your 1080 Ti results:
> 
>   Number of GPUs detected: 1
>   #0: NVIDIA GeForce GTX 1080 Ti, compute cap.: 6.1, ECC: no, stat: comp
>   [ ... ]
>   starting mdrun [ ... ]
>   50000 steps,    210.0 ps.
>   Writing final coordinates.
>                  Core t (s)   Wall t (s)        (%)
>          Time:      112.643       14.080      800.0
>                    (ns/day)    (hour/ns)
>   Performance:     1288.622        0.019
> 
> 
> 
> And this is the run on an almost identical system with
> a GTX 1080 Ti (Palit Super-Jetstream):
> 
>   Number of GPUs detected: 1
>   #0: NVIDIA GeForce GTX 1080, compute cap.: 6.1, ECC:  no, stat: comp
>   [ ... ]
>   starting mdrun 'VILLIN in water'
>   50000 steps,    210.0 ps.
>   Writing final coordinates.
>                  Core t (s)   Wall t (s)        (%)
>          Time:      147.840       18.480      800.0
>                    (ns/day)    (hour/ns)
>   Performance:      981.841        0.024
> 
> 
> On the (small) villin benchmark, the 1080 Ti would be
> about 13% faster than the GTX 1080. The raw SP float power
> of the 1080 Ti is about 40% higher than the 1080 (11,340
> GFLOPS vs 8,228 GFLOPS) which means a faster processor
> could possibly help here.
> 
> 
> BTW: I have a large membrane benchmark (DPPC/water, 1.2M atoms,
> 35x35x13A³ box) which runs at about 3 ns/d on the GTX 1080 with
> Parrinello-Rahman semiisotropic coupling, if you'd like to torture
> your box I can provide it ;)
> 
> Regards
> 
> M.
> 
> -- 
> Gromacs Users mailing list
> 
> * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting!
> 
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> 
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-request at gromacs.org.