The goal of the Tensor Algebra Computations Over Manycore Architectures (TACOMA) project is to design a high-performance tensor contractions package for heterogeneous systems with multicore CPUs and GPU accelerators. TACOMA will target the acceleration of a number of important applications, including high-order finite element method (FEM) simulations, machine learning, big data analytics, multi-physics, and more.
The TACOMA project is a collaboration between UTK/ICL, Lawrence Livermore National Lab (LLNL), the University of Paris-Sud, and Inria to develop software infrastructure for tensor contractions to, in part, expedite the porting of packages like BLAST to next generation 150+ petaflop/s DOE machines (e.g., Sierra, LLNL’s IBM Power9 machine). The TACOMA effort is multidisciplinary, incorporating linear algebra, languages, code generation and optimizations, domain science, and application-specific numerical algorithms. TACOMA is funded until March 2017, at which time it will be superseded by the CEED project.
In Collaboration With
- Lawrence Livermore National Laboratory
- University of Paris-Sud