Re: [AMBER] CUDA driver insufficient

From: wei zhang <>
Date: Tue, 26 Nov 2013 15:35:55 -0800 (PST)

Dear All,
As described in my previous Emails,
I have complied the cuda and it runs fine at the Dell machine.   However,  to  make PMEMD_cuda work, I had to
fist uninstall the original Dell-nvidia package.  I actually had to uninstall it using the window’s
graphic interface :  System setting/
add/removing application ( rpm –e doen’t work, shown GPU not installed…).   However, by doing so,  I lost the login windows.  To make things worse, after rebooting the
server, PMEMD_cuda stopped working,  even
after I recompile it.
During the recompiling, some
abnormal returns and warnings were shown as below:
Advisory: Loop was not unrolled, unexpected control flow construct
Advisory: Loop was not unrolled, unexpected control flow construct
Advisory: Loop was not unrolled, cannot deduce loop trip count
Advisory: Loop was not unrolled, cannot deduce loop trip count
ptxas warning : Too many threads
per SM specified for entry
will be ignored”
 make test.cuda hangs on:
cd nmropt/gb/angle/ &&
./Run.nmropt_1angle_gb  SPFP
 Similar hanging happed using as
my originally compile PMEMD_cuda after the rebooting,  ( PMEMD_cuda running  but doesn’t output, seems like a dead circle…)
 Would very much appreciate you
Meanwhile,  any hints on how can I get the login window
back?  Tried to reinstall the
original-Dell Nvidia drive back, but shown confliction with current  driver ( same thing happened when install the
latest NVIDIA driver, that’s why I  had
to remove the Dell-driver to use PMEMD_cuda). 
Thanks a lot!


 From: wei zhang <>
To: wei zhang <>; Ross Walker <>; AMBER Mailing List <>; AMBER Mailing List <>
Sent: Wednesday, November 20, 2013 9:09 AM
Subject: Re: [AMBER] CUDA driver insufficient

Here are some updates about the CUDA driver.
Driver 331.20 works fine for tesla C2075.
PMEMD testing resutls on the machine (  Dual six-core Dell-Precision R5500n):
1np: 1X
2np: 1.74X
4np: 3.65X
8np: 6.36X
10np: 7.38X
cuda: 23.39X
Best regards,

From: wei zhang <>
To: Ross Walker <>; AMBER Mailing List <>
Sent: Tuesday, November 19, 2013 9:19 AM
Subject: Re: [AMBER] CUDA driver insufficient

Thanks! Ross,
I gpt the drives from the NVDIA website, below is the link:
(Seems somehow I used advanced search, but 304 listed as recommandered driver still can be misleading..)
Best regards,

From: Ross Walker <>
To: wei zhang <>; AMBER Mailing List <>
Sent: Monday, November 18, 2013 10:26 PM
Subject: Re: [AMBER] CUDA driver insufficient

Hi Wei,

This happens when the driver you have loaded does not match the version of
cuda, and importantly the cuda runtime library. I.e. NVIDIA has minimum
driver requirements for each version of CUDA. The 304 driver sounds
ancient to me. Not sure where you are finding that one from. I doubt it
works for anything beyond cuda 5.0 which is likely the source of your

Note AMBER also has a minimum requirement with the latest GPUs (700
series) of 319.60 (recommended) or 325.15 (which is actually older than
319.60). I have not tried the 331 branch which is brand new and completely
untested. Hopefully the fixes in the 319 and 325 trees got propogated into
the 331 tree but you will likely be the first person trying it.

Where did you see 304 being recommended? - I'll let NVIDIA know to fix
their website.

All the best

On 11/18/13 3:35 PM, "wei zhang" <> wrote:

>Thanks! Jason.
>Using 331.20 indeed solved the problem.
>A bit supprise though..  ( both 331.20 and 304.1160 are the latest
>version released on Nov6, 2013, while the recommanded one did not work
>for some reason)
>Best regards,

> From: Jason Swails <>
>To: wei zhang <>; AMBER Mailing List
>Sent: Monday, November 18, 2013 4:32 PM
>Subject: Re: [AMBER] CUDA driver insufficient
>On Mon, Nov 18, 2013 at 5:08 PM, wei zhang <> wrote:
>Dear All,
>>I am try to compile AMBER12 on a Dell-Precision R5500n machine.
>>It have a single tesla C2075 GPU.
>>Everything seems OK untill "make test.cuda" retuned a bunch of errors
>>indicating" CUDA driver version is insufficient for CUDA runtime version"
>>I used the latest CUDA-5.5 toolkits, and the NVIDIA recommadered driver:
>>Linux x64 (AMD64/EM64T) Display Driver (version: 304.1160).
>>Could anyone provide some hints? should I use a older CUDA?
>No, you should use a later driver.  There were issues with drivers before
>325.15 for certain hardware, so there was a check put in to make sure
>that was the version driver being used.
>I believe the C2075 is fine with the 304.1160 drivers (maybe).  If you
>can't upgrade the driver for whatever reason, then you can always comment
>out the driver version check in gpu.cpp and recompile (but I would
>suggest trying to get the updated driver, first).
>Good luck,
>Jason M. Swails
>Rutgers University
>Postdoctoral Researcher
>AMBER mailing list
AMBER mailing list
AMBER mailing list
AMBER mailing list
Received on Tue Nov 26 2013 - 16:00:03 PST
Custom Search