Porting CUDA-Based Molecular Dynamics Algorithms to AMD ROCm Platform Using HIP Framework: Performance Analysis

The use of graphics processing units (GPU) in computer data processing tasks has long ceased to be an unusual event. However, now GPU computing is nearly synonymous with CUDA, a proprietary framework for developing applications for Nvidia's GPU devices. It provides comprehensive documentation and excellent development tools. Meanwhile, the main competitor of Nvidia in the market for the production of GPU devices, the AMD company is developing its own Radeon Open Compute (ROCm) platform that features an application programming interface compatible with CUDA. The primary objective of this work is to investigate whether ROCm provides a worthy alternative to CUDA in the field of GPU computing. The work has two sub-objectives: the description of the programmers experience investigation during porting classical molecular dynamics algorithms from CUDA to ROCm platform and performance benchmarking of initial and resulting programs on GPU devices with modern architectures (Pascal, Vega10, Vega20).