Characteristics of Nvidia CUDA and AMD ROCm Platforms Affecting Performance Portability

9 Jul 2021, 12:15
15m
403 or Online - https://jinr.webex.com/jinr/j.php?MTID=mf93df38c8fbed9d0bbaae27765fc1b0f

403 or Online - https://jinr.webex.com/jinr/j.php?MTID=mf93df38c8fbed9d0bbaae27765fc1b0f

https://jinr.webex.com/jinr/j.php?MTID=mf93df38c8fbed9d0bbaae27765fc1b0f
Sectional reports 5. High Performance Computing HPC

Speaker

Vsevolod Nikolskiy (HSE)

Description

The development and popularization of the AMD ROCm platform with HIP technology allows one to create code that is not locked to a specific vendor maintaining a high level of performance. A lot of legacy but still supported codes is originally written in CUDA, and now it is getting ROCm HIP support as well. In a recent paper [1], the performance of popular molecular dynamics packages with GPU support was discussed in detail. The research includes the LAMMPS package providing backends for CUDA, OpenCL, and HIP. Based on this package, we can compare and define in detail the platform properties and performance impact of real parallel code. Differences can be found in the characteristics of the target hardware, the operation of the software environment and drivers, and even in the logic of the application code itself. In continuation of the study, the work of computational GPU kernels in the application using several MPI processes for each GPU is considered.

1. Kondratyuk N, Nikolskiy V, Pavlov D, Stegailov V. GPU-accelerated molecular dynamics: State-of-art software performance and porting from Nvidia CUDA to AMD HIP. The International Journal of High Performance Computing Applications. April 2021. doi:10.1177/10943420211008288

Primary author

Presentation materials

There are no materials yet.