Hi Marek,
You are comparing the 'wrong' thing. K20 / K20X are HPC cards, this means
SLOWWWWWW (but green ;-) ). Thus you should really compare the K20X
performance with the M2090 which is the other HPC card. So there you have:
DHFR/NVE
M2090 = 43.74 NS/day
K20X = 89.13 ns/day
So that's a more than doubling in performance which is pretty much in line
with Moore's law.
The GTX Titan you should compare against the GTX680 as they are both
gaming cards. And thus FASSSTTTT (and hot!). There you'll be looking at
54.46ns/day and I expect about 115ns/day or so for the GTX Titan board so
yeap pretty much Moore's law as well.
So I am not sure what numbers you were looking at to say the performance
difference was not big. Perhaps you looked at the TRPCage numbers which is
the first plot on the page and GB? That won't tell you much since it is
only 304 atoms! There you are pretty much at the parallel limit so it
doesn't benefit from the extra cores in the new cards. I might move the
implicit solvent benchmarks to the bottom of the page to avoid this
confusing people. For comparison you should really look at the explicit
solvent benchmarks on the amber page:
http://ambermd.org/gpus/benchmarks.htm#Benchmarks
With regards to your question about optimization for K20. There is
potential for about another 30% performance improvement or so but this is
quote a bit of work and has to be balanced against adding more features.
All the best
Ross
On 2/19/13 8:55 AM, "Marek Maly" <marek.maly.ujep.cz> wrote:
>Hi Filip and all,
>
>I just obtained offer to test K20 here:
>
>http://www.nvidia.co.uk/object/k20-gpu-test-drive-uk.html
>
>so it forced me to check actual Amber benchmarks where K20/K20x
>results are already present so one may compare
>their performance e.g. to GTX680.
>
>http://ambermd.org/gpus/benchmarks.htm
>
>
>as anybody can see the increase of performance
>is not so overwhelming.
>
>So in this context I do not understand Filip's
>
>"much much faster" .
>
>much much faster THAN WHAT ???
>
>Perhaps not than K20/K20x if yes why ?
>
>Anyway I would be grateful for any comments regarding so small
>performance increase of the new architectures K20/K20x
>comparing to GTX608.
>
>I am not sure but if I remember well there was some
>opinions that the new Kepler will be about
>5x times faster (maybe not than GTX680 which is K104 based
>but than fermi based GTX580 which is also not true)
>
>Is there problem, that the Amber code is still not fully
>optimised for K20 and the actual patch 14 just allowed
>to use Amber on these new generation GPUs but it will
>need some more code improvements in order to use all
>the advantages of the new Kepler architecture ?
>
>
> Best wishes,
>
> Marek
>
>
>
>
>
>
>Dne Tue, 19 Feb 2013 17:40:12 +0100 filip fratev <filipfratev.yahoo.com>
>napsal/-a:
>
>> Hi Ross and all,
>>
>> After a lot of rumors the GTX Titan card was
>> released today. For the first time on any consumer-level NVIDIA card,
>> double precision (FP64)
>> performance is uncapped. That means 1/3 FP32 performance, or
>> roughly 1.3TFLOPS theoretical FP64 performance. Thus this card is very
>> similar
>> (same for Amber use) to Tesla K20x, but costs 1000$ and will be much
>>much
>> faster! I suppose that Titan will break the 100+ ns threshold on JAC
>> test with
>> Amber 12 and have no patience to see some test results!
>> All the best,
>> Filip
>> _______________________________________________
>> AMBER mailing list
>> AMBER.ambermd.org
>> http://lists.ambermd.org/mailman/listinfo/amber
>
>
>--
>Tato zpráva byla vytvořena převratným poštovním klientem Opery:
>http://www.opera.com/mail/
>
>_______________________________________________
>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 Feb 19 2013 - 10:30:02 PST
