Optimizing Use of Different Types of Memory for FPGAs in High Performance Computing

Abstract

Accelerators such as Field Programmable Gate Arrays (FPGAs) are increasingly used in high performance computing, and the problems they are applied to process larger and larger amounts of data. FPGA manufacturers have added new types of memory on chip to help ease the memory bottleneck; however, the burden is on the designer to determine how data is allocated to different memory types. We study the use of ultraRAM for a graph application running on Amazon Web Services (AWS) that generates a large amount of intermediate data that is not subsequently accessed sequentially. We investigate different algorithms for mapping data to ultraRAM. Our results show that use of ultraRAM can speed up overall application run time by a factor of 3 or more. Maximizing the amount of ultraRAM used produces the best results, and as problem size grows, judiciously assigning data to ultraRAM vs. DDR results in better performance.