Hello Filip,
Thank you for sharing your experience with the GTX 470. Many of us are hoping, like you, that Nvidia will release a workaround to this issue. We know they are working on it.
Cheers,
Sergio
Sergio Aragon
Professor of Chemistry
SFSU
-----Original Message-----
From: filip fratev [mailto:filipfratev.yahoo.com]
Sent: Saturday, December 25, 2010 3:09 AM
To: AMBER Mailing List
Subject: Re: [AMBER] GTX470 and Amber 11 Cuda tests
In fact I didn’t test JAC PME (64x64x64) by some longer calculations and when I tried now they crashed *randomly* identically to the other GTX 470 users:
Error: the launch timed out and was terminated launching kernel kPMEGetGridWeights
cudaFree GpuBuffer::Deallocate failed the launch timed out and was terminated
For sure it is not a temperature issue. I tested the memory and it was ok. I hope that the Amber and Nvidia teams will solve the problem.
I was able just to make the benchmark: 22.15ns/day (SPDP).
Regards,
Filip
--- On Sat, 12/25/10, filip fratev <filipfratev.yahoo.com> wrote:
> From: filip fratev <filipfratev.yahoo.com>
> Subject: [AMBER] GTX470 and Amber 11 Cuda tests
> To: amber.ambermd.org
> Date: Saturday, December 25, 2010, 9:14 AM
> Dear Amber users,
> First of all Marry Christmas and Happy New Year!!
>
> I did some tests with Asus GTX470 and Amber 11 (bug fixes
> 1-11 applied) and Phenom II X6 1090T.4.0GHz, Fan Speed 100%
> by Coolbits option set to 5 and the Temps 67C Max GPU during
> the tests and MB 33C. Unfortunately by some reasons I used
> Cuda version 3.1 not 3.2.
> Here are some of the results using the Amber benchmark
> tests for CUDA:
>
> Explicit Solvent
> 1) JAC-NTV: 21.92 ns/day
> 2) JAC-NTP: 18.97 ns/day
> 3) JAC-NTP 6x1090T: 5.63 ns/day
> 4) FactorIX NVE: 5.56 ns/day
> 5) FactorIX NTP: 3.68 ns/day
>
> GB
> 1) TRPCage: 380.14 ns/day
> 2) Myoglobin: 48.82 ns/day
>
> As you can see form the above examples and those on the
> Amber web, GTX470 performed better than even C2050 for the
> small systems. The overcklocking is still no possible before
> Nvidia to issue some drives for that but we could expect up
> to 20% increase of the performance after overcklock, i.e.
> similar performance to the new GTX570. Thus we hope that the
> bug issue with those cards will be solved by Amber and
> Nvidia teams!
>
> Interestingly, my simulations did not crash after short MD
> tests as it was reported previously by some other GTX470
> users:
> http://archive.ambermd.org/201009/0198.html
> I run few 500ps to 2 ns simulations on JAC and also the
> example from MMPBSA and everything was ok.
>
> Regards,
> Filip
>
>
>
>
>
> _______________________________________________
> 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 Sat Dec 25 2010 - 11:00:04 PST
