Software:tinkergpu

Tinker9 GPU Molecular Dynamics

Tinker9 is the next generation of Tinker software with GPU accelaration. As of 2021, Ren lab switched to Tinker9 for GPU simulations, which is similar/consistent to Tinker CPU in terms of usage/setup, is faster than OpenMM for AMOEBA and some fixed charged force fields, and has all our new development for AMOEBA+.

For single molecule or cluster without PBC, please CPU code. Use openmp-threads to speed up if necessary

The setup and key files used by Tinker CPU (on this wiki page) can be directly applied to Tinker9.

Source code

https://github.com/TinkerTools/tinker9

Free to academic and nonprofit org

Compile Tinker9

Please read Tinker9 Github site for how to setup the environmental variables and compile Tinker9 code (https://github.com/TinkerTools/tinker9). Successful builds on various hardware and cuda version can also be found at https://github.com/TinkerTools/tinker9/discussions/121.

Note that Tinker9 required "matching" canonical Tinker 8, which should be automated by CMAKE script now.https://github.com/TinkerTools/tinker9/blob/master/README.md  

As of May 23, 2022, Ren lab clusters equiped with GPU cards all have cuda version 11 with compatitable cuda driver. Here is a script file to compile tinker9

#!/bin/bash

# build on bme-sugar
# Chengwen Liu

rm -fr *.sh src* *Make* cmake-* tinker9

export CUDAHOME=/usr/local/cuda-11.2
export CUDACXX=$CUDAHOME/bin/nvcc
export FC=/usr/bin/gfortran
export CXX=/usr/bin/g++
export ACC=/home/liuchw/shared/nvidia/hpc_sdk/Linux_x86_64/21.1/compilers/bin/nvc++
export opt=release
export host=0
export prec=m
export compute_capability=60,70,75,80
export cuda_dir=$CUDAHOME
export CMAKEHOME=/home/liuchw/shared/cmake3.21/bin/

$CMAKEHOME/cmake ..

make -j

This is to say, after downloading Tinker9, make a build directory in tinker9 home directory. Then running the above should do everything.

Example

Example setup, xyz and key files for protein simulations: https://github.com/TinkerTools/tinker9/blob/master/example/

Recommend to use RESPA integrator with 2fs time step and write out less frequent (e.g. every 2 ps). On RTX3070, you should be able to achieve ~40ns/day for DHFR. Use MonteCarlo or Langvin piston for pressure control in NPT More details see Tutorials

Manual

https://tinkerdoc.readthedocs.io/en/latest/

Run Script

Here is a simple script to run molecular dynamics

#!/bin/bash

# set the GPU device
export CUDA_DEVICE_ORDER=PCI_BUS_ID
export CUDA_VISIBLE_DEVICES=0 # set to 1,2, etc to specify the one you want

# add tinker9 executables to PATH
export TINKER9=/home/liuchw/Softwares/tinkers/Tinker9/2205/build_cuda11
export PATH=$PATH:$TINKER9

# option 1 to run dynamic
nohup dynamic9 your.xyz -k your.key 1000 2.0 2.0 4 298.15 1.0 >your.log 2>&1 &

# option 2 to run dynamic
nohup tinker9 dynamic your.xyz -k your.key 1000 2.0 2.0 4 298.15 1.0 >your.log 2>&1 &

To use Tinker 9 on RTX4090 or A6000, node15x & node 16x

#!/bin/bash
export TINKER9=/home/pren/tinker9/tinker9_xeon/build/
export CUDA_DEVICE_ORDER=PCI_BUS_ID
export CUDA_VISIBLE_DEVICES=0 # device number if there are multiple GPU cards
$TINKER9/tinker9 dynamic bench7.xyz -k bench7.key 5000 2.0 400.0 4 298.15 1.0 N > out &

The above also tested on RTX 30x0 with older xeon processors (tacc node, see "lscpu")