Maximize Performance and Portability of HPC Apps with HPC SDK v21.11

The latest NVIDIA HPC SDK includes a variety of tools to maximize developer productivity, as well as the performance and portability of HPC applications.   

Today, NVIDIA announced the upcoming HPC SDK 21.11 release with new Library enhancements. This software will be available free of charge in the coming weeks.

The NVIDIA HPC SDK is a comprehensive suite of compilers and libraries for high-performance computing development. It includes a wide variety of tools proven to maximize developer productivity, as well as the performance and portability of HPC applications.   

The HPC SDK and its components are updated numerous times per year with new features, performance advancements, and other enhancements. 

What’s new

This 21.11 release will include updates to HPC C++/Fortran compiler support and the developer environment, as well as new multinode mulitGPU library capabilities. 

Compiler, build systems, and other enhancements

Introduced last year with version 20.11, the NVFORTRAN compiler automatically parallelizes code written using the DO CONCURRENT standard language feature as described in this post

New in version 21.11, the programmer can use the REDUCE clause as described in the current working draft of the ISO Fortran Standard to perform reduction operations, a requirement of many scientific algorithms.

Starting with the 21.11 release, the HPC Compilers now support the –gcc-toolchain option, similarly to the clang-based compilers. This is provided in addition to the existing rc-file method of specifying nondefault GNU Compiler Collection (GCC) versions. The HPC Compilers leverage open source GCC libraries for things like common system operations and C++ standard library support. 

Sometimes, a developer needs a different version of the GCC toolchain than the system default. Now 21.11 has both command line and file-based ways of making that specification. In addition to –gcc-toolchain, the 21.11 HPC Compilers add several GCC-compatible command line flags for specifying x86-64 target architecture details.

The 21.11 release now includes two new Fortran modules to integrate with NVIDIA libraries, Fortran applications maximize the benefit from NVIDIA platforms and Fortran developers be as productive as possible. HPC applications written in Fortran can directly use cufftX—a highly optimized multi-GPU FFT library from NVIDIA. It also enables easier use with the NVIDIA Tools Extension Library (NVTX) for performance and profiling studies with Nsight.

Version 21.11 will ship with CMake config files that define CMake targets for the various components of the HPC SDK. This offers application packagers and developers a more seamless code integration with the NVIDIA HPC SDK.

New multinode, multiGPU Math Libraries

HPC SDK version 21.11 will include the first of our upcoming multinode, multiGPU Math Library functionality, cuSOLVERMp. Initial functionality will include Cholesky and LU Decomposition, with and without pivoting. Future releases will include LU, with multiple RHS. 

Figure 1. Gain roughly 2x performance improvement against the current state-of-the-art multi-node software. 

Learn more about:

Leave a Reply

Your email address will not be published. Required fields are marked *