Hardware Acceleration of HPC Computational Flow Dynamics using HBM-enabled FPGAs

Abstract

Scientific computing is at the core of many High-Performance Computing (HPC) applications, including computational flow dynamics. Because of the uttermost importance to simulate increasingly larger computational models, hardware acceleration is receiving increased attention due to its potential to maximize the performance of scientific computing. A Field-Programmable Gate Array (FPGA) is a reconfigurable hardware accelerator that is fully customizable in terms of computational resources and memory storage requirements of an application during its lifetime. Therefore, it is an ideal candidate to accelerate scientific computing applications because of the possibility to fully customize the memory hierarchy important in irregular applications such as iterative linear solvers found in scientific libraries. In this paper, we study the potential of using FPGA in HPC because of the rapid advances in reconfigurable hardware, such as the increase in on-chip memory size, increasing number of logic cells, and the integration of High-Bandwidth Memories on board. To perform this study, we first propose a novel ILU0 preconditioner tightly integrated with a BiCGStab solver kernel designed using a mixture of High-Level Synthesis (HLS) and Register-Transfer Level (RTL) hand-coded design. Second, we integrate the developed preconditioned iterative solver in Flow from the Open Porous Media (OPM) project, a state-of-the-art open-source reservoir simulator. Finally, we perform a thorough evaluation of the FPGA solver kernel in both standalone mode and integrated into the reservoir simulator that includes all the on-chip URAM and BRAM, on-board High-Bandwidth Memory (HBM), and off-chip CPU memory data transfers required in a complex simulator software such as OPM's Flow. We evaluate the performance on the Norne field, a real-world case reservoir model using a grid with more than 10 5 cells and using 3 unknowns per cell, and we find that the FPGA is on par with the CPU execution and 3 times slower than the GPU implementation when comparing only the kernel executions.