Dear Ross,
Thank you very much for your helpful suggestions.
For the option 2, we have implemented the two protein systems; one is our
system which has successfully been manipulated by GPU amber in our other
older machine that equips with amber18&ambertool18, ubuntu 14.04LTS, CUDA
7.5 and a gpu GTX780. The other system is DHFR protein taken from the "test"
package in amber18. For both systems, the RTX2080Ti machine failed to
implement calculations with GPU amber. The error message obtained from each
system is the same: GpuBuffer:Download failure.
So, we think that we can rule out the possibility of the option 2.
For the option 1, we examined the gpu-check commands such as "nvidia-smi
(-L)" and "nvcc --version" according to various instruction in ubuntu
forums, and these seem that the gpu works correctly. However, as per your
suggestion, we further have checked the gpu itself by using sample programs
stored in CUDA directory
(/usr/local/cuda-10.2/bin/cuda-install-samples-10.2.sh). After the commend
"make", we encountered a fatal error during compiling sample programs by
nvcc.
This may suggest that there is something wrong with our RTX2080Ti card
mechanically.
We are going to send our machine back to the manufacture for checking the
hardware.
If the manufacture could not identify the causes, as a plan B, we will wait
the release of amber20 to ensure the possibility of the option 3.
Thank you for your kind help.
Best wishes,
Keigo
-----Original Message-----
From: Ross Walker <ross.rosswalker.co.uk>
Sent: Sunday, January 19, 2020 4:32 AM
To: AMBER Mailing List <amber.ambermd.org>
Subject: Re: [AMBER] GPU-AMBER18 installation on RTX2080Ti
Hi Keigo,
I have not yet tested AMBER 19 with CUDA 10.2 myself although I'd expect it
to work. CUDA 10.1 works fine. The warnings during the compilation are
innocuous. You can ignore those. However the GpuBuffer:Download failure is a
fatal error. To confirm do you see this when running the GPU test cases or
when you are running a simulation of a protein of your own? This sort of
error can typically occur due to three things, 1) A malfunctioning GPU, bad
memory, flaky GPU core etc, 2) An unstable simulation, clashes of atoms,
insufficiently minimized etc, 3) A subtle error coming from the combination
of compilers, system library versions used etc.
If you see this when running the test cases then that rules out option 2.
All the best
Ross
> On Jan 18, 2020, at 05:29, <ke.gohda.camm-kansai.org>
<ke.gohda.camm-kansai.org> wrote:
>
> Dear AMBER users,
>
>
>
> Recently, we have purchased a machine equipped with a single RTX2080Ti
> GPU for simulation in microsecond scale.
>
> For setting up GPGPU circumstance on the machine, we have chosen the
> combination of system/application as follows:
>
> - amber18 & ambertool 19
>
> - ubuntu 18.04.3 LTS (kernel 5.0.0-37)
>
> - nvidia driver 440.33.01 (latest as of this moment)
>
> - cuda 10.2.89
>
>
>
> The installation and testing for CPU version of amber have
> successfully been completed without any errors.
>
> However, we have encountered fatal errors in testing of GPU version of
> amber.
>
>
>
> During the installation of GPU amber, no fatal errors were reported.
> However, a bunch of warring messages were taken place from gputypes.h
> and gpuBuffer.h, for instance. I am not an expert on CUDA language,
> but it may suggest that the communication between the GPU and amber
> will not work appropriately.
>
>
>
> In fact, the fatal error message from testing of GPU amber is
> "cudaMemcpy GpuBuffer::Download failed an illegal memory access was
> encountered", suggesting confirmation of the assumption above.
>
>
>
> Then, we are planning to change the combination of the
> system/application to make GPU amber be operative.
>
> Because we believe that many of amber users have successfully
> implemented RTX2080Ti GPUs in their machines, we want to mimic the
> successful combination on GPU circumstance. Actually, pre-configured
> amber18 machines with RTX2080Ti are available through AMBER HP.
>
>
>
> We would very much appreciate it if you could let us know the
> specification of your RTX2080Ti machines with amber18 on the following
system/application:
>
> - version of ubuntu
>
> - kernel version of the ubuntu, if appropriate
>
> - version of nvidia driver
>
> - version of cuda
>
>
>
> Thank you for your kind help in advance.
>
>
>
> Best wishes,
>
> Keigo Gohda, PhD
>
> Computer-aided Molecular Modeling Research Center, Kansai
> (CAMM-Kansai)
>
>
>
> _______________________________________________
> 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 Mon Jan 20 2020 - 04:00:02 PST
