A floating-point accumulator for FPGA-based high performance computing applications

A floating-point accumulator for FPGA-based high performance computing applications is proposed and evaluated. Compared to previous work, our accumulator uses a fixed size circuit, and can reduce an arbitrary number of input sets of varying sizes without requiring prior knowledge of the bounds of summands. In this paper, we describe how the adder accumulator operator can be heavily pipelined to achieve a high clock speed when mapped to FPGA technology, while still maintaining the original input ordering. Our experimental results show that our accumulator design is very competitive with previous efforts in terms of FPGA resource usage and clock frequency, making it an ideal building block for large-scale sparse matrix computations as implemented in FPGA-based high performance computing systems.