Re: [AMBER] Are benchmarks/tests on NVidia RTX 2080 Super available?

From: Pietro Amodeo <pamodeo.icb.cnr.it>
Date: Sat, 29 Feb 2020 21:26:38 +0100

Thank you to Abhilash an Ross for their replies and info.

As for the RTX2080 Super performances, in principle the possible effects
of core and memory overclocking over 2080 and the overall balancing
between lower cores/mem. bandwith and higher core/mem.clock vs. 2080Ti
weren't so evident to me. Thus I found very useful the raw number from
Ross, showing for 2080 Super performances on NTP simulations variable
from 80% (for large systems) to 90% (for small systems) of the 2080Ti.
Since the former costs about 60% of the corresponding 2080Ti model, it
could represent a viable alternative, provided that 8GB vs. 11Gb is not
a serious issue and that enough PCI-E 3.0 16x slots are available to
host the extra cards that can be purchased with the same overall budget.


The number of slots, or better, the card density, represents another
important issue with recent NVidia GPUs. While already a serious problem
(at least) from GTX980Ti, with the RTX2080 series it worsened both
quantitatively and qualitatively, since, in addition to requiring
increasingly better thermal solution, NVidia changed their reference
design, that now isn't any more suited for server cases. Other cooling
solutions (included those blowing air out of the case) typically require
either low densities (<=2cards/case) or a careful choice of the case
and/or possible addition of fans/ducts and removal of back plates.
However, all these solutions must be carefully tested at maximum load
(possibly for both CPUs and GPUs). In this view, in my lab the systems
we are going to assemble will be water-cooled, a solution requiring some
extra money and work and the programming of an efficient thermal
management, but easier to design and ensuring lower noise/wear (this
latter all-external to GPUs, if waterblock solution are adopted) and
considerably better thermal performances.

Best,

Pietro

> From: Abhilash J <md.scfbio.gmail.com>
> Date: Thu, 27 Feb 2020 12:48:41 -0500
>
> Hi Ross,
>
> Thanks. I will try get the data center folks make some modifications and
> see what performance gains i can get.
>
> Regards
>
> Abhilash
>
> On Thu, Feb 27, 2020 at 12:45 PM Ross Walker <ross.rosswalker.co.uk> wrote:
>
> > Hi Abhilash,
> >
> > Since it is in a data center I highly recommend asking your data center
> > folk to set the case fans in the bios to be maximum speed. This should help
> > a lot.
> >
> > All the best
> > Ross
> >
> > > On Feb 27, 2020, at 12:42 PM, Abhilash J <md.scfbio.gmail.com> wrote:
> > >
> > > Hi Ross,
> > >
> > > Thanks for the insight. I am using CUDA Version: 10.2. nvidia smi
> > > output is attached below.
> > > As you say the AMBER 16 is not optimized for RTX2XXX cards. I will try
> > > to work the problem out by trying to get AMBER 18 there.
> > > The new RTX 2080 Super cards are housed in a data center they have 4
> > > cards per node. The MD test was carried out with 3 cards being idle and 1
> > > card running the simulation, so i hoped heating would not be an issue.
> > > My local workstation has 4 cards (1080 Ti) and the output has been
> > > around 36 ns/day (+- 1-2ns/day) in all different combinations.
> > >
> > > ======================================
> > >
> > >
> > +-----------------------------------------------------------------------------+
> > > | NVIDIA-SMI 440.31 Driver Version: 440.31 CUDA Version:
> > 10.2 |
> > >
> > |-------------------------------+----------------------+----------------------+
> > > | GPU Name Persistence-M| Bus-Id Disp.A | Volatile
> > Uncorr. ECC |
> > > | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util
> > Compute M. |
> > >
> > |===============================+======================+======================|
> > > | 0 GeForce RTX 208... Off | 00000000:3B:00.0 Off |
> > N/A |
> > > | 24% 24C P8 8W / 250W | 275MiB / 7981MiB | 0%
> > Default |
> > >
> > +-------------------------------+----------------------+----------------------+
> > > | 1 GeForce RTX 208... Off | 00000000:5E:00.0 Off |
> > N/A |
> > > | 25% 24C P8 10W / 250W | 12MiB / 7982MiB | 0%
> > Default |
> > >
> > +-------------------------------+----------------------+----------------------+
> > > | 2 GeForce RTX 208... Off | 00000000:AF:00.0 Off |
> > N/A |
> > > | 24% 23C P8 4W / 250W | 12MiB / 7982MiB | 0%
> > Default |
> > >
> > +-------------------------------+----------------------+----------------------+
> > > | 3 GeForce RTX 208... Off | 00000000:D8:00.0 Off |
> > N/A |
> > > | 25% 22C P8 13W / 250W | 12MiB / 7982MiB | 0%
> > Default |
> > >
> > +-------------------------------+----------------------+----------------------+
> > >
> > >
> > +-----------------------------------------------------------------------------+
> > > | Processes: GPU
> > Memory |
> > > | GPU PID Type Process name
> > Usage |
> > >
> > |=============================================================================|
> > > | 0 3127 C python
> > 115MiB |
> > > | 0 5715 G /usr/bin/X
> > 84MiB |
> > > | 0 6307 G /usr/bin/gnome-shell
> > 70MiB |
> > >
> > +-----------------------------------------------------------------------------+
> > > =======================================
> > >
> > >
> > >
> > >
> > >
> > > On Thu, Feb 27, 2020 at 12:14 PM Ross Walker <ross.rosswalker.co.uk>
> > wrote:
> > >
> > >> Hi Abhilash
> > >>
> > >> It's most likely the difference in AMBER version although the difference
> > >> is bigger than I'd expect. The RTX architecture did not exist when
> > AMBER 16
> > >> came out so there is no optimization for it. Also what CUDA compiler
> > >> version are you using? You ideally need to use 10.0 or later for RTX2XXX
> > >> cards.
> > >>
> > >> A stock RTX2080 should be a little bit faster than a stock 1080TI. The
> > >> RTX2080 Super is really just an overclocked RTX2080, in my experience
> > it's
> > >> about 5 to 10% faster than a stock 2080 depending on the simulation
> > >> parameters. The small difference in 2080 vs 2080 Super performance is
> > why
> > >> I never bothered to update the benchmarks on the AMBER website.
> > >>
> > >> Note the other possibility is that you are seeing throttling of the
> > >> RTX2080 due to heat issues. If you have more than 1 GPU in a box this
> > can
> > >> be a big issue with RTX cards. NVIDIA's reference design is not great
> > for
> > >> cooling. The ducted systems that blow air out of the back of the case
> > are
> > >> better but if you have multiple cards next to each other you have to
> > remove
> > >> the backplate on the cards so there is room for airflow. As a quick
> > test I
> > >> would recommend going into the machine bios and setting the case fan
> > speeds
> > >> to maximum (it may sound like a jet engine). Then rerun the benchmark.
> > That
> > >> will tell you if it's a thermal throttling issue.
> > >>
> > >> You can monitor the GPU temperature during a run using nvidia-smi. I
> > >> forget the exact number but it's something like 80C that the RTX2080 is
> > >> capped at so if the temperature pegs at that it means the clock speed is
> > >> likely being reduced to prevent overheating.
> > >>
> > >> All the best
> > >> Ross
> > >>
> > >>
> > >>> On Feb 27, 2020, at 11:06 AM, Abhilash J <md.scfbio.gmail.com> wrote:
> > >>>
> > >>> Hi Everyone,
> > >>>
> > >>> I tried my hands on RTX 2080 Super today. I used AMBER 16. I got 17
> > >>> ns/day for a system of 172K atoms.
> > >>> This was quite disappointing as i get 36 ns/day from my 1080 ti cards
> > >>> using AMBER 18 using the same 172K atoms system.
> > >>> I don't know if it is the variation in the GPU cards or the one lower
> > >>> AMBER version responsible.
> > >>> Both machines were built by Exxact corporation.
> > >>> I hope some developers can weigh in on this.
> > >>>
> > >>> Regards
> > >>>
> > >>> Abhilash
> > >>>
> > >>>
> > >>>
> > >>>
> > >>>
> > >>> On Sat, Feb 22, 2020 at 8:05 AM Pietro Amodeo <pamodeo.icb.cnr.it>
> > >> wrote:
> > >>>
> > >>>> Hi,
> > >>>>
> > >>>> since NVidia RTX 3XXX cards are not expected before Q4 2020, and
> > rumored
> > >>>> RTX 2080 Super Ti models have not been confirmed, RTX 2080 Super,
> > >>>> costing between 3/5 and 2/3 of Ti, could represent a good interim
> > >>>> solution, especially because both all-in-one and integrated-block
> > >>>> water-cooled cards based on this chip exist (a potentially critical
> > >>>> issue for such a clock-boosted solution).
> > >>>>
> > >>>> In this view, info about RTX 2080 Super performances with AMBER 18
> > (and,
> > >>>> possibly, 20) would allow a proper cost/benefit ratio comparison
> > between
> > >>>> Super and Ti versions.
> > >>>>
> > >>>> Thank you in advance for any info.
> > >>>>
> > >>>> Best,
> > >>>>
> > >>>> Pietro
> > >>>> _______________________________________________
> > >>>> 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
> >
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber

-- 
Dr Pietro Amodeo, PhD
Primo Ricercatore CNR
Istituto di Chimica Biomolecolare del CNR
Comprensorio "A. Olivetti", Edificio 70
Via Campi Flegrei 34
I-80078 Pozzuoli (Napoli) - Italy
Phone    +39-0818675072
Mobile    +39-3469711696
Fax        +39-0818041770
Email    pamodeo.icb.cnr.it
StOrMoDB Molecular Database: https://stormodb.na.icb.cnr.it/stormodb/
(EN [1])
MoSeF Science Communication Network: http://mosef.na.cnr.it (IT [2]/EN
[3]) 
Links:
------
[1] https://stormodb.na.icb.cnr.it/stormodb/
[2] http://mosef.na.cnr.it/it
[3] http://mosef.na.cnr.it/en
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Sat Feb 29 2020 - 12:30:02 PST
Custom Search