Memory Optimized Architecture for Efficient Gauss-Jordan Matrix Inversion

This paper presents a new architecture for efficient Gauss-Jordan matrix inversion algorithm on reconfigurable hardware platforms. The results show that currently available re- configurable computing technology can easily achieve significantly higher floating-point performance than high-end CPUs, running state-of-the-art routines for large matrices operations. For common reconfigurable systems, where the FPGAs are directly coupled to the on-board memory, the achievable performance scales directly with the number of realizable simultaneous memory accesses. A new dedicated reconfigurable architecture is proposed and analysed and the results show a performance improvement of 2x over the previous implementation, using only half of the memory and half of the floating-point units. Benchmarking against Matlab, which features high performance matrix inversion routines, shows that a 100 MHz FPGA can easily surpass the performance of 3,2 GHz Intel Pentium IV processors. This is possible having only 5 double-port memory banks or 9 single-port memory banks connected to the FPGA.