It's also clocked significantly higher and that's what's making the real difference. I do worry a bit about long-term calculations on them though without really stable power and cooling. One thing I would do for your C2050 is turn off ECC. That's good for 10%...
DPDP is a very interesting conundrum because the kernels spill into global memory like crazy and that mitigates the potential performance gain from 4x the double-precision. This could probably be tweaked though.
-----Original Message-----
From: Ross Walker [mailto:ross.rosswalker.co.uk]
Sent: Monday, June 14, 2010 13:31
To: 'AMBER Mailing List'
Subject: Re: [AMBER] pmemd.cuda DPDP performance on GTX480 vs C2050?
Hi Sasha,
This is what I suspected. The GT480 is quicker than the C2050 for SPSP and
SPDP since it has more cores 480 vs 448 on the C2050. The interesting part
is that DPDP is not as slow as I thought it would be on the 480. The SPDP
performance is also quite a bit better than I expected over the C2050.
I just wonder if one can cool it enough to prevent it catching fire. ;-)
Note, the SPDP and, if you are conservative, the DPDP are the only results
you should consider. The SPSP mode we have considerable evidence that it
produces a number of artefacts in the simulation.
All the best
Ross
> -----Original Message-----
> From: Sasha Buzko [mailto:obuzko.ucla.edu]
> Sent: Monday, June 14, 2010 12:10 PM
> To: AMBER Mailing List
> Subject: Re: [AMBER] pmemd.cuda DPDP performance on GTX480 vs C2050?
>
> Hi Ross,
> I ran the Factor IX benchmark with your input file on a GTX480.
> Surprisingly, it's not significantly slower than your C2050 test, at
> least with this system.
> With your input, it completed in 956.84 s total time, or about 1.806
> ns/day (DPDP mode). The SPSP version finished in 239.8 s (7.206
> ns/day).
> And the default SPDP version took 257.41 s or 6.713 ns/day.
>
> For the sake of completeness, the GTX480 is on a desktop with dual
> dual-core Xeon 5160 (3.00 GHz) and 8 GB of RAM.
>
> Sasha
>
>
> Ross Walker wrote:
> > Hi Sahsa
> >
> >
> >> has anyone compared performance of pmemd.cuda in DPDP mode on GTX480
> >> and
> >> C2050?
> >> Tesla cards are supposed to have a better double precision
> performance,
> >> but how much better in the event of an all-double pmemd code?
> >> Thanks for any info
> >>
> >
> > I don't have a GTX480 so have not been able to compare the
> differences,
> > however, here are the numbers for C2050, I'd be interested to see the
> > difference.
> >
> > FACTOR - IX
> >
> > NVE 128x64x64 FFT
> > Typical Production MD NVE with
> > GOOD energy conservation.
> > &cntrl
> > ntx=5, irest=1,
> > ntc=2, ntf=2, tol=0.000001,
> > nstlim=10000,
> > ntpr=1000, ntwx=1000,
> > ntwr=10000,
> > dt=0.002, cut=8.,
> > ntt=0, ntb=1, ntp=0,
> > ioutfm=1,
> > /
> > &ewald
> > dsum_tol=0.000001,nfft1=128,nfft2=64,nfft3=64,
> >
> > /
> >
> >
> > SPSP SPDP (Default) DPDP
> > Specs Time (s) NS/day Time (s) NS/day Time (s)
> > NS/day
> > 8 proc N/A N/A 1034 1.671179884
> > GTX295 395.46 4.369594902 525.6 3.287671233 2469.18
> 0.699827473
> > C1060 366.25 4.718088737 492.84 3.506208912 2354.73
> 0.733842097
> > C2050 311.34 5.550202351 333.04 5.188565938 706.98
> 2.44419927
> >
> >
> >
> >
> > Myoglobin = 2492 atoms
> >
> > &cntrl
> > imin=0,irest=1,ntx=5,
> > nstlim=10000,dt=0.002,ntb=0,
> > ntf=2,ntc=2,tol=0.000001,
> > ntpr=1000, ntwx=1000, ntwr=50000,
> >
> > cut=9999.0, rgbmax=15.0,
> > igb=1,ntt=0,nscm=0,
> > /
> >
> > SPSP SPDP (Default) DPDP
> > Specs Time (s) NS/day Time (s) NS/day Time (s)
> > NS/day
> > 8 proc N/A N/A 395.73 4.3666136
> > GTX295 39.62 43.61433619 64.99 26.58870595 3.0 Compiler
bug
> > #VALUE!
> > C1060 37.79 45.72638264 61.94 27.89796577 3.0 Compiler
bug
> > #VALUE!
> > C2050 31.39 55.04937878 34.6 49.94219653 155.7
11.0982659
> >
> > All the best
> > Ross
> >
> >
> >
> >
> > _______________________________________________
> > 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
-----------------------------------------------------------------------------------
This email message is for the sole use of the intended recipient(s) and may contain
confidential information. Any unauthorized review, use, disclosure or distribution
is prohibited. If you are not the intended recipient, please contact the sender by
reply email and destroy all copies of the original message.
-----------------------------------------------------------------------------------
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Mon Jun 14 2010 - 14:30:03 PDT
