Hi Andrew
The card you.mention looks like it is reference <cough>founders </cough> design already. I recommend only ever using such cards. The overstocked multi fan systems are fine for games but unreliable for cuda apps and the things like the 3 fan designs they have are terrible in multi-task configuration. It also helps to use server cases with ducted cooling - e.g. 2 U boxes like this
http://exxactcorp.com/index.php/solution/solu_detail/201
This clocks down a whole lot less than desktop boxes.
You can set cool bits in your xorg.conf file and set the fans to 100% this stops things clocking down as well in multi gpu configuration but your system will sound like a bunch of jet engines in a box.
Didn't know about the exclusive mode not working. Yet another big. I'll check it an file with nvidia. They rushed out the gtx1080 3 months before it was ready so it is still massively experimental. At least they fixed the fact it was getting incorrect numerical results (4th time now /fail) quickly with the 367.27 driver. But the fact you have to get a release candidate of the Cuda compiler to use the card at launch is just freaking ridiculous.
All the bestRoss
-------- Original message --------From: "Beavil, Andrew" <andrew.beavil.kcl.ac.uk> Date: 6/23/16 14:49 (GMT-07:00) To: amber.ambermd.org Subject: [AMBER] Amber16 + GTX1080 - working well
I have two GTX1080s (EVGA ACX3.0) running amber and the key is to ensure you apply update 6, which came out last week.
Without update 6, you will get the error : "cudaMemcpyToSymbol: SetSim copy to cSim failed invalid device symbol"
I’m using : Amber16/Cuda8.0RC/Nvidia driver 367.27 on Centos 7
it’s probably best to use ‘make distclean’ before running configure as that cleans out old half compiled junk and forces the updates to be applied.
The one remaining issue I’m having is that the 367.27 driver (which is needed for the GTX1080s) doesn’t seem to support ‘exclusive process’ mode i.e., 'nvidia-smi -c 3’ - they two cards work fine if I use ‘export CUDA_VISIBLE_DEVICES=...'
In terms of performance, for a system similar in size to the DHFR benchmark, running NTP MD with 2fs steps & 10A cutoff I get
~110 ns/day on a GTX980
~135ns/day on GTX980ti
~170ns/day on GTX1080
performance drops slightly if both cards are running at the same time because the physically lower card heats the higher one somewhat.
Although the performance boost is welcome, the thing I like most is that you can control the power usage of these cards and if I set the GTX1080 to 120watts, I get ~150ns/day (similar but better than the GTX980ti) for only half the power consumption of the GTX980ti.
[it makes for a cooler and quieter office!]
[the cards run at 1900MHz flat out ~165watts or 1500MHz 90watts - 120watts seems a sensible compromise and leads to ~1800Mhz]
Does anyone have experience of the founders edition cards and are they better at pushing heat out of the back of the system ?
ie less heating of neighbouring cards ?
Best wishes,
Andrew Beavil
-------------------------------------------------------------------------
Dr. Andrew J. Beavil, Phone: (+44) (0)20 78488064
Senior Lecturer in Molecular Biophysics Fax : (+44) (0)20 78486410
MRC & Asthma UK Centre in Allergic Mechanisms of Asthma
King's College London,
The Randall Division,
New Hunt's House, Email: andrew.beavil.kcl.ac.uk<mailto:andrew.beavil.kcl.ac.uk>
Guy's Campus,
London Bridge,
London. SE1 1UL
-------------------------------------------------------------------------
Date: Thu, 23 Jun 2016 10:35:33 -0400
From: Zack Scholl <zns.duke.edu<mailto:zns.duke.edu>>
Subject: Re: [AMBER] AMBER support for gtx1080
To: AMBER Mailing List <amber.ambermd.org<mailto:amber.ambermd.org>>
Message-ID:
<
CAH7kaX=e+C684h47UJh-99x624rs7pO-NVvhdW_Kd-9qEbG-Zg.mail.gmail.com<mailto:
CAH7kaX=e+C684h47UJh-99x624rs7pO-NVvhdW_Kd-9qEbG-Zg.mail.gmail.com>>
Content-Type: text/plain; charset=UTF-8
In our lab we upgraded to GTX1080 from GTX980Ti and we are indeed seeing
about 30% improvement. Make sure you have Amber16 and CUDA-8.0 (the
developer version). Also be aware you might have to get new drivers (we're
using RHEL 7 and the card wouldn't load on the system until we updated
them). You can download CUDA8.0 and get the new drivers here:
https://developer.nvidia.com/cuda-toolkit.
Currently I'm a little worried something is wrong with the GTX1080 with
AMBER16. Though it initially worked for me, it has since failed every
simulation saying simply "cudaMemcpyToSymbol: SetSim copy to cSim failed
invalid device symbol". Reinstalling Amber16 has not succeeded in fixing
this error. I'd be curious if you or anyone else has run into this. I will
try to reinstall CUDA8.0 to see if that is a possible fix, but I can't
think of a way to reset the system beyond that (except reinstalling the OS
I suppose!). If I keep running into this I'd probably suggest getting a
GTX980Ti until the GTX1080 system is more stable.
--
_______________________________________________
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 Fri Jun 24 2016 - 06:30:03 PDT
