Can GPUs be used to speed up MPI computations

[Anders_S_1]

Hi,

I am running FP-heavy computations using Fortran and MPI under Windows. It is not possible to

use OpenMP. Is it possible to offload FP calculations to a GPU to get a speed up? Do GPUs offer double precision calculations?

Best regards

Anders S

[Barbara_P_Intel]

For a Fortran application to offload to an Intel GPU, OpenMP directives are required. MPI optimizations around offload are related to managing where a given rank offloads.

The availability of double precision FP calculations varies by the Intel GPU. For example, the GPUs targeted for gaming have single precision FP.

[Anders_S_1]

Hi Barbara,

Thanks for your swift reply!

Which Intel GPUs offer double precision?

If I understood you right, work can be offloaded from each MPI rank using OpenMP directives. Is there any rule of thumb when offload to a GPU will result in a speedup as a function of size of the MPI task?

Is there any sample code or example on offload of a double precision workload to a GPU?

Is it possible to evaluate a simple example in a "sandbox"?

Best regards

Anders S

[Barbara_P_Intel · Accepted solution]

Here's some info to help you get started.

Three Quick, Practical Examples of OpenMP Offload to GPUs (video)

https://www.intel.com/content/www/us/en/developer/videos/three-quick-practical-examples-openmp-offload-gpus.html

Run HPC Applications on CPUs & GPUs with Xe Architecture Using Intel® C++ & Intel® Fortran Compilers with OpenMP* (video)

Run HPC Applications on CPUs & GPUs with Xe Architecture Using Intel C++ and Intel Fortran Compilers with OpenMP

Basic understanding of GPU architecture

https://www.intel.com/content/www/us/en/develop/documentation/oneapi-gpu-optimization-guide/top/xe-arch.html

oneAPI GPU Optimization Guide

https://www.intel.com/content/www/us/en/develop/documentation/oneapi-gpu-optimization-guide/top.html

As far as a "sandbox", use Intel DevCloud for oneAPI.

[Anders_S_1]

Hi Barbara,

As for the sandbox, will it be possible to evaluate offload to the MAX 1100 GPU (or similar) in the near future?

Best regards

Anders S

[Barbara_P_Intel]

I don't know the roadmap for DevCloud. There is a Community Forum for DevCloud questions.