Thanks to Ross and Pratul for helping out here. I'll just emphasize that
Amber's GPU code is really insensitive to the CPU. The speed of that
processor affects less than 1% of the calculation, so any CPU core will be
able to handle the communications and kernel launches. The only exceptions
would be if you are doing special cases of GaMD or NEB which involve hybrid
CPU / GPU calculations, in which case your run speed will be partly
dependent on the speed of the CPU.
I'll post more to the list as the story of my 300W RTX cards develops.
Dave
On Mon, Dec 17, 2018 at 8:55 PM Ross Walker <ross.rosswalker.co.uk> wrote:
> Hi Rui,
>
> Note Gold 5115 CPUs are overkill for GPU AMBER unless you also plan to run
> a lot of CPU based calculations. You can likely back this off to Silver
> 4114 CPUs and save yourself about $1400 or a so a node.
>
> In terms of the GPUs either option is good assuming the vendor supplying
> them is properly testing them to make sure they give correct numerical
> results and that the cooling is sufficient that the cards do not thottle
> during benchmarking. You can use the Amber benchmark suite from the GPU
> page to test this. That will run in each GPU in turn and then on all GPUs
> at once. In both cases you should see identical performance for all GPUs
> whether they are being used individually or all at the same time. Note to
> achieve this with RTX2080TI in 4x or 8x configs, where there is not space
> between the GPUs required developing a custom cooling solution. This is
> what Exxact had to do for their AMBER systems (
> https://www.exxactcorp.com/AMBER-Certified-MD-Systems <
> https://www.exxactcorp.com/AMBER-Certified-MD-Systems>) so if you are
> using a different vendor you should ask them what their cooling solution
> is, what the card base model they are using is, and if they can guarantee
> there won't be throttling due to heat.
>
> Note if the 4 GPU system is really a 4.5U 8 GPU box with GPUs spaced by 2
> PCI slots in each case then you should be okay but if it is a 2U x 4GPU box
> you will have the same issues as the 8 GPU system.
>
> Hope that helps. Let me know if you want help speccing anything up further
> - disk, memory etc.
>
> All the best
> Ross
>
> > On Dec 17, 2018, at 19:46, Rui Sun <ruisun.hawaii.edu> wrote:
> >
> > Thank you for the quick response, Dave.
> >
> > If I may bother you with another question, the options that I have right
> > now are:
> > #1: *4* units of RTX 2080 + 2 units of Intel Gold 5115 per node ($16,000)
> > #2: *8* units of RTX 2080 + 2 units of Intel Gold 5115 per node ($26,000)
> >
> > Apparently, the 8-unit node will be more cost-effective but do you think
> I
> > might have a cooling issue?
> >
> > Best,
> > Rui
> >
> >
> > *Rui Sun*
> > Assistant Professor
> > Department of Chemistry
> > University of Hawaii at Manoa
> > Bilger 245B
> > 2545 McCarthy Mall
> > Honolulu, HI 96822-2275
> > Phone: (808) 956-3207
> >
> >
> > On Fri, Dec 14, 2018 at 1:13 PM David Cerutti <dscerutti.gmail.com>
> wrote:
> >
> >> The RTX-2080Ti is performing very well, but be careful about the
> cooling!
> >> I want to release a patch and I know what fixes to make, but I still
> don't
> >> have a good test platform as the card in my new workstation is getting
> up
> >> to 88*C (it'll shut down for safety purposes at 89). The card is also
> not
> >> putting out the performance that RTX-2080Tis in Ross's machines, which
> seem
> >> to have better cooling, are able to do. This is a 300W card--and while
> 300
> >> versus 250W may not seem like a big deal consider the excess heating in
> a
> >> confined volume of the same size inside a system of the same size with
> the
> >> same fans. About like you if you started eating two extra candy bars a
> >> day--the calories would add up fast. So the benchmark numbers on the
> >> website are genuine, and the GB portion may even come up to speed with
> >> Volta more once we retune those kernels for Turing, but understand that
> >> this horse needs lots of water.
> >>
> >> Dave
> >>
> >>
> >> On Fri, Dec 14, 2018 at 5:54 PM Rui Sun <ruisun.hawaii.edu> wrote:
> >>
> >>> Dear AMBER Users,
> >>>
> >>> I was wondering if I could get some information on the performance of
> >>> AMBER18 on the recently-released RTX 2080Ti. How is it comparing to
> Titan
> >>> V?
> >>>
> >>> Currently, I am considering buying a few GPU nodes and I am currently
> >>> debating between the following two configurations:
> >>> #1: 4 units of Titan V + 2 units of Intel Gold 5115 per node
> >>> #2: 8 units of RTX 2080 + 2 units of Intel Gold 5115 per node
> >>>
> >>> Thank you so much,
> >>> Rui
> >>>
> >>>
> >>> *Rui Sun*
> >>> Assistant Professor
> >>> Department of Chemistry
> >>> University of Hawaii at Manoa
> >>> Bilger 245B
> >>> 2545 McCarthy Mall
> >>> Honolulu, HI 96822-2275
> >>> Phone: (808) 956-3207
> >>> _______________________________________________
> >>> 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 Dec 17 2018 - 20:00:02 PST
