Update on AMD/HIP support in Amber22

(updated version created on 3 January, 2023)

We are pleased to announce the availability of support for AMD GPU hardware in pmemd version 22. Users should understand that this involves a lot of new code, and that we are looking for feedback on any problems that arise. Please start with short test runs, and check the outputs carefully before undertaking long simultions.

This update/patch can be used if you already have Amber22 installed. Do the following:

1. Download https://drive.google.com/file/d/1_CBKH84KYhO85WTKZYaZCj50fotWA1cU/view?usp=sharing
2. Navigate to your amber22_src folder.
3. Untar the file you downloaded
4. Follow the instructions in the compile_with_hip.sh file.

You should probably do the above in a fresh copy of the amber22_src folder, just to be sure nothing bad happens to your current Amber22 installation. Please report successes and failures to the amber mailing list, so that your experience helps others.

More information from AMD is available here.

Introduction

The fastest academic GPU MD simulation engine, pmemd.cuda, is written and maintained by researchers in the Amber community; see literature references below. Principal current and past developers include:

• David Cerutti, overseeing major code renovations, performance enhancements, and maintenance of the general MD engine
• Taisung Lee, co-author of the thermodynamic integration and free energy feature extensions
• Daniel Mermelstein, co-author of the thermodynamic integration and free free energy feature extensions
• Charles Lin, co-author of the GPU NMR restraint code, thermodynamic integration and free energy extensions
• Perri Needham, co-author of the GPU NMR restraint code
• Delaram Goreishi, author of Nudged Elastic Band methods in CUDA and Fortran
• Ross Walker, project and QA lead, author of the first CUDA extensions for the original

  pmemd

  Fortran program and developer of the mixed precision models

The state of the code is also buoyed by the generous support of Ke Li, Peng Wang, Duncan Poole and Mark Berger (technology engineers and alliance managers) at NVIDIA Corporation, and Andrew Nelson, Nick Chen and Mike Chen at Exxact Corporation.

Since the advent of GPU accelerated simulations in Amber11, the engine has taken on new features, quality control mechanisms, and algorithms. While the inherently parallel GPU architecture does not permit the verbose error checking and reporting that the CPU code contains, we actively monitor user feedback and engage a set of built-in debugging functions to help us understand any issues that arise. Hundreds of labs and companies all over the world use the latest Amber22 GPU simulation engine.

The code supports serial as well as parallel GPU simulations, but from Pascal (2016) onward, the benefit of running a simulation, with the exception of REMD based simulations, on two or more GPUs is marginal. On the latest Voltai and Turing architectures our algorithms cannot scale to multiple GPUs. We therefore recommend executing independent simulations on separate GPUs in most cases. A key design feature of the GPU code is that the entirety of the molecular dynamics calculation is performed on the GPU. This means that only one CPU core is needed to drive a simulation and a server full of four or eight GPUs can run one independent simulation per card without loss of performance provided that there are at least the same number of free CPU cores available as GPUs in use. (Most commodity CPU chips have at least four cores.) The fact that GPU performance is unaffected by CPU performance means that any CPU compiler (the open source GNU C and Fortran compilers are adequate) will deliver comparable results with Amber's premier engine, and sets Amber apart from other molecular dynamics codes. Another key feature of this design choice is that it means low cost CPUs can be used which coupled with custom designed precision models and bitwise reproducibility use to validate consumer cards gives AMBER unrivaled performance per dollar.

Literature references

The initial Amber implementation papers, covering implicit and explicit solvents:

• Andreas W. Goetz; Mark J. Williamson; Dong Xu; Duncan Poole; Scott Le Grand;   & Ross C. Walker* "Routine microsecond molecular dynamics simulations with AMBER - Part I: Generalized Born", J. Chem. Theory Comput., 2012, 8 (5), pp 1542-1555, DOI: 10.1021/ct200909j
• Romelia Salomon-Ferrer; Andreas W. Goetz; Duncan Poole; Scott Le Grand;  & Ross C. Walker* "Routine microsecond molecular dynamics simulations with AMBER - Part II: Particle Mesh Ewald", J. Chem. Theory Comput., 2013, 9 (9), pp 3878-3888. DOI: 10.1021/ct400314y
• Scott Le Grand; Andreas W. Goetz; & Ross C. Walker* "SPFP: Speed without compromise - a mixed precision model for GPU accelerated molecular dynamics simulations.", Comp. Phys. Comm, 2013, 184, pp374-380, DOI: 10.1016/j.cpc.2012.09.022

More recent thermodynamic integration capabilities are described here:

• Tai-Sung Lee, Dan Mermelstein, Charles Lin, Scott LeGrand, Timothy J. Giese, Adrian Roitberg, David A. Case, Ross C. Walker* & Darrin M. York*, "GPU-accelerated molecular dynamics and free energy methods in Amber18: performance enhancements and new features", J. Chem. Inf. Mod., 58:2043-2050 2018. 10.1021/acs.jcim.8b00462
• Tai-Sung Lee, Yuan Hu, Brad Sherborne, Zhuyan Guo, & Darrin M. York*, "Toward Fast and Accurate Binding Affinity Prediction with pmemdGTI: An Efficient Implementation of GPU-Accelerated Thermodynamic Integration", J. Chem. Theory Comput., 2017, 13, pp 3077–3084, DOI: 10.1021/acs.jctc.7b00102
• Daniel J. Mermelstein, Charles Lin, Gard Nelson, Rachael Kretsch, J. Andrew McCammon, & Ross C. Walker*, "Fast, Flexible and Efficient GPU Accelerated Binding Free Energy Calculations within the AMBER Molecular Dynamics Package", J. Comp. Chem., 2018, DOI: 10.1002/jcc.25187