ICL Research Profile

Overview

In the Ginkgo project, we develop high performance numerical linear algebra functionality reflecting the parallelism of modern HPC platforms. The focus is on GPU-accelerated systems, and Ginkgo can currently be used on AMD GPUs, Intel GPUs, and NVIDIA GPUs using backends written in the respective vendor languages. Ginkgo features a variety of Krylov solvers, sparse direct solvers, sophisticated preconditioners exposing fine-grain parallelism, including incomplete factorizations, incomplete sparse approximate inverses, and algebraic multigrid technology, mixed precision algorithms, and preconditioned batched iterative solvers.

Ginkgo is implemented in modern C++ and is used by several popular simulation frameworks, including MFEM, SUNDIALS, deal.ii, HyTeg, openCARP, XGC, NEK5000, and OpenFOAM. Scalability of the Ginkgo backend to up to thousands of GPUs has been demonstrated on the Frontier and Perlmutter Supercomputers. The Ginkgo library is open source under the 3-clause BSD license. In the development of Ginkgo, we aim for industry-level code quality standards, including Continuous Integration (CI) and Continuous Benchmarking (CB), comprehensive unit test coverage, and fulfilling the community policies of the extreme-scale Scientific Software Development Kit (xSDK) and the Extreme Scale Scientific Software Stack (E4S).