Just ran the benchmarks on my own 2080Ti, CUDA 10.0:
Report of all timings (ns/day):
System Class CPU (1) CPU (4) GPU 1
------------------------------- ---------------- -------- -------- --------
JAC_production_NVE_4fs PME (Standard) 676.71
JAC_production_NPT_4fs PME (Standard) 600.70
Cellulose_production_NVE_4fs PME (Standard) 51.38
Cellulose_production_NPT_4fs PME (Standard) 49.28
FactorIX_production_NVE_4fs PME (Standard) 227.52
FactorIX_production_NPT_4fs PME (Standard) 212.09
STMV_production_NVE_4fs PME (Standard) 18.30
STMV_production_NPT_4fs PME (Standard) 17.52
JAC_production_NVE_4fs PME (Optimized) 723.65
JAC_production_NPT_4fs PME (Optimized) 661.62
Cellulose_production_NVE_4fs PME (Optimized) 53.93
Cellulose_production_NPT_4fs PME (Optimized) 51.88
FactorIX_production_NVE_4fs PME (Optimized) 247.81
FactorIX_production_NPT_4fs PME (Optimized) 229.95
STMV_production_NVE_4fs PME (Optimized) 19.43
STMV_production_NPT_4fs PME (Optimized) 18.63
TRPCage GB 2139.36
myoglobin GB 1056.37
nucleosome GB 23.22
I saw nothing unusual about the large STMV system. Do you have the latest
benchmark suite? It was updated in the winter.
Dave
On Tue, Jul 23, 2019 at 2:53 PM David Cerutti <dscerutti.gmail.com> wrote:
> We've compiled with 10.0 and 10.1 for RTS-2080Ti and had no such
> problems. Not sure about the driver numbers. Can you try this: go into
> your config.h and change "-gencode arch=compute_70,code=sm_70" to read _75
> rather than _70. I don't think that's the problem but give it a try.
>
> Dave
>
>
> On Tue, Jul 23, 2019 at 12:57 PM Giovanni Grazioso <
> giovanni.grazioso.unimi.it> wrote:
>
>> Il 23/07/19 18:36, David Cerutti ha scritto:
>> > I'm not familiar with any problems specific to RTX-2080Ti like this.
>> The
>> > benchmark is failing, too? How are you running the benchmark, exactly?
>>
>> yes, I have run the runBenchmarks.sh.
>>
>> > Does the benchmark fail immediately, or after 60-90 seconds? (It should
>> > take about three minutes to run.)
>>
>> The MD simulations of all systems in the test package were concluded
>> without any problem, with benchamarks similar to the one reported on the
>> website and after about 30 min.
>>
>> The problem was encountered when finally the "STMV_production_NPT_4fs"
>> was investigated.
>>
>> The simulations started but at step 9000/10000 several **** appear in
>> the mdout. Similar problems were found when the time_step is reduced to
>> 2fs.
>>
>> In order to understand if there are structural defects on my
>> GPU/workstation, I'd like to know which are the characteristics
>> (compilers, Linux version, cudatoolkit version, and nVidia driver) of
>> the workstation usually implementing RTX2080Ti for pmemd.cuda
>> calculations.
>>
>> Thanks for you reply, Dave.
>>
>> Giovanni
>>
>>
>> >
>> > Dave
>> >
>> >
>> > On Tue, Jul 23, 2019 at 8:56 AM Giovanni Grazioso <
>> > giovanni.grazioso.unimi.it> wrote:
>> >
>> >> Dear Amber Users,
>> >>
>> >> I have some problems with RTX2080Ti/MSI on my workstation
>> (Centos7,
>> >> Cudatoolkit10.1, kernel3.10, nVidia driver 430.34, 32GB RAM,
>> 4xIntel-i7).
>> >>
>> >> I have successfully compiled amber18/amber19tools/cuda by gnu compiler.
>> >> Then, testing pmemd.cuda by the Amber18_BenchmarkSuite, when the
>> >> "STMV_production_NPT_4fs" was simulated, this message appeared on the
>> >> terminal:
>> >>
>> >> " Error: an illegal memory access was encountered launching kernel
>> >> kScaleVelocities",
>> >>
>> >> the md0_mX.out is:
>> >>
>> >> | MC Barostat: Decreasing size of volume moves
>> >> check COM velocity, temp: NaN NaN(Removed)
>> >>
>> >> NSTEP = 9000 TIME(PS) = 5157.003 TEMP(K) = NaN
>> >> PRESS = 0.0
>> >> Etot = NaN EKtot = NaN EPtot =
>> >> **************
>> >> BOND = 0.0000 ANGLE = 21082677.2413 DIHED =
>> >> 0.0000
>> >> 1-4 NB = ************** 1-4 EEL = ************** VDWAALS =
>> >> -3595.1791
>> >> EELEC = -19399252.5829 EHBOND = 0.0000 RESTRAINT
>> >> = 0.0000
>> >> EKCMT = 0.0000 VIRIAL = 0.0000 VOLUME =
>> >> 11082169.1330
>> >> Density
>> >> = 1.0032
>> >>
>> >>
>> >> Unfortunately, trying to perform NPT/MD on different systems (the ones
>> >> of my interest), after 5 min of simulations (NPT_2fs) the calculations
>> >> crash and the energy contributions in mdout looked as before.
>> >>
>> >> NVE calculations proceed normally (without any ***) but the run is
>> >> blocked without any message in the terminal or in the mdout.
>> >>
>> >> By using GTX1080Ti, NVE and NPT calculations proceed without any
>> >> problems. I've tried to use different version of cudatoolkit and nvidia
>> >> driver but the problems remained.
>> >>
>> >>
>> >> Can somebody suggest where the problem is?
>> >>
>> >> If someone uses RTX20280Ti, could specify which version of nVidia
>> driver
>> >> and cudatoolkit works correctly?
>> >>
>> >> Thanks in advance
>> >>
>> >> Giovanni
>> >>
>> >>
>> >> --
>> >>
>> >>
>> >>
>> >>
>> >>
>> >> *****************************************************************
>> >> Giovanni GRAZIOSO
>> >> Associate Professor
>> >> Dipartimento di Scienze Farmaceutiche - DISFARM
>> >> Università degli Studi di Milano
>> >> Via L. Mangiagalli 25 - 20133 Milano (Italy)
>> >> Office: +039/02503-19352 - Lab: +039/02503-19351
>> >> Fax: +039/02503-19359 - E-mail: giovanni.grazioso.unimi.it
>> >> *****************************************************************
>> >>
>> >>
>> >>
>> >>
>> >> _______________________________________________
>> >> 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
>>
>>
>> --
>>
>>
>>
>>
>>
>> *****************************************************************
>> Giovanni GRAZIOSO
>> Associate Professor
>> Dipartimento di Scienze Farmaceutiche - DISFARM
>> Università degli Studi di Milano
>> Via L. Mangiagalli 25 - 20133 Milano (Italy)
>> Office: +039/02503-19352 - Lab: +039/02503-19351
>> Fax: +039/02503-19359 - E-mail: giovanni.grazioso.unimi.it
>> *****************************************************************
>>
>>
>>
>>
>>
>> _______________________________________________
>> 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 Tue Jul 23 2019 - 15:30:03 PDT
