Computational Linear Algebra on FPGAs
Research topic and goals
As the Dennard scaling ended, hardware manufacturers and application scientists have turned to non-traditional architectures to serve the need for compute power while keeping the power-budget low. For mainstream CPU and GPU architectures, there exists a large collection of libraries providing basic linear algebra building blocks (vector and matrix operations) and numerical algorithms (linear solvers, preconditioners, etc.) that are developed either by the hardware manufacturers (“vendor libraries”) or the academic community. For exotic architectures like FPGAs, in many cases the user community is yet too small to drive the development on basic linear algebra subroutines or numerical algorithms. This, in turn, reduces the potential use of these architectures in scientific computations, in particular computational science.
In this project, JSC researchers and ICL researchers will collaborate to develop and expand the existing linear algebra libraries and functionality. The project is expected to accelerate the development of linear algebra functionality for FPGAs, and ensure the functionality hits the architecture-specific performance limits.
Results for 2023/2024
Albert spent 3 months at ICL and worked extensively on LU Decomposition Kernel for Xilinx FPGA, specifically, U55C. This kernel was integrated with the public release of OOPS library under the OPTIMA Project(https://optima-hpc.eu/). Furthermore, the results of this kernel we part of Deliverable 5.5(https://optima-hpc.eu/wp-content/uploads/2023/12/D5.5.pdf). Besides the deliverable, we also gained expertise in performance measurements for linear algebra routines and that is also reflected in Deliverable 6.2 of the OPTIMA Project.
Visits and meetings
- 3-month stay of Albert Njoroge Kahira (JSC) at the University of Tennessee
Impact and publications
The collaboration helped the successful completion of the OPTIMA HPC project and was highlighted as a KPI in the project.
Future plans
None.