The latest NVIDIA HPC SDK includes a variety of tools to maximize developer productivity, as well as the performance and portability of HPC applications.
Overview
NVIDIA announced the upcoming release of HPC SDK 21.11, which includes significant enhancements aimed at improving the performance and portability of high-performance computing applications. Key updates include new compiler features, multinode multiGPU library capabilities, and improved integration with NVIDIA libraries.
What You'll Learn
1
How to utilize the REDUCE clause in Fortran for scientific algorithms
2
How to specify a nondefault GCC toolchain using the –gcc-toolchain option
3
Why the new cuSOLVERMp library can enhance multiGPU performance
Prerequisites & Requirements
- Familiarity with Fortran programming and high-performance computing concepts
- Access to NVIDIA HPC SDK and compatible hardware
Key Questions Answered
What new features are included in HPC SDK v21.11?
HPC SDK v21.11 introduces updates to the HPC C++/Fortran compiler support, a new REDUCE clause for Fortran, and multinode multiGPU library capabilities, including cuSOLVERMp for advanced mathematical operations.
How does the NVFORTRAN compiler improve code parallelization?
The NVFORTRAN compiler automatically parallelizes code using the DO CONCURRENT standard language feature, enhancing performance for scientific applications by simplifying parallel code development.
What is cuSOLVERMp and what functionality does it provide?
cuSOLVERMp is a new library in HPC SDK v21.11 that offers multinode, multiGPU functionality for mathematical operations, initially supporting Cholesky and LU Decomposition, which are critical for many scientific computations.
Key Statistics & Figures
Performance improvement
roughly 2x
cuSOLVERMp offers this performance improvement against the current state-of-the-art multi-node software.
Technologies & Tools
Software Suite
Nvidia Hpc SDK
A comprehensive suite of compilers and libraries for high-performance computing development.
Library
Cusolvermp
Provides multinode, multiGPU functionality for mathematical operations.
Key Actionable Insights
1Leverage the new REDUCE clause in Fortran to optimize your scientific algorithms, as it allows for efficient reduction operations that are essential in many computational tasks.This feature can significantly enhance the performance of applications that rely on reduction operations, making it a valuable tool for developers working in scientific computing.
2Utilize the –gcc-toolchain option to easily switch between different GCC versions, ensuring compatibility and performance optimization for your HPC applications.This flexibility is crucial for developers who need to work with specific GCC versions for their projects, enhancing productivity and reducing setup time.
3Explore the capabilities of cuSOLVERMp for multiGPU operations to achieve performance improvements in your applications, potentially doubling the efficiency compared to existing solutions.As HPC applications increasingly require efficient multiGPU processing, utilizing cuSOLVERMp can provide a significant edge in performance.
Common Pitfalls
1
Failing to specify the correct GCC toolchain can lead to compatibility issues and hinder application performance.
Developers should ensure they are using the appropriate GCC version for their projects, as mismatches can cause build failures or runtime errors.
Related Concepts
High-performance Computing
Fortran Programming
Multigpu Processing
Nvidia Libraries