FPGA-based Custom Computing Accelerator for Computational Fluid Dynamics based on Lattice Boltzmann Method
FPGA-based Custom Computing Accelerator for Computational Fluid Dynamics based on Lattice Boltzmann Method
复制标题
基于 FPGA 的定制计算加速器，用于基于格子玻尔兹曼方法的计算流体动力学
DOI：
10.1007/978-3-319-10626-7_16
复制
发表时间：
2014
期刊：
影响因子：
--
通讯作者：
Kentaro Sano
中科院分区：
文献类型：
--
作者：
Kentaro Sano;Hayato Suzuki;Ryo Ito;Tomohiro Ueno;and Satoru Yamamoto;Kentaro Sano
提出了一种用于流体动力学仿真定制计算的紧密耦合FPGA集群，并通过原型实现对其性能进行了评价。为了在FPGA加速器众多的情况下实现可扩展和高效的计算，我们提出了一种加速器域网络（ADN），通过直接连接FPGA实现低延迟和高速的数据传输。我们描述了一个带有四个fpga的原型集群节点的实现，以及它们的片上框架，用于高速数据流和计算。在性能评估中，我们证明了我们的网格玻尔兹曼方法（LBM）用于流体动力学计算的定制计算机器利用了时间和空间并行性，并且随着fpga的数量很好地扩展了性能。结果，我们用四个fpga实现了73.0 GFlop/s峰值性能的98.8%。
This paper presents a tightly-coupled FPGA cluster for custom computing of fluid dynamics simulation, and evaluates its performance with prototype implementation. For scalable and efficient computation with a lot of FPGA accelerators, we propose an accelerator-domain network (ADN) that brings low-latency and high-speed data transfer by directly connecting FPGAs. We describe implementation of a prototype cluster node with four FPGAs, and their on-chip framework for high-speed data streaming and computing. In performance evaluation, we demonstrate that our custom computing machine for fluid dynamics computation with the lattice-Boltzmann method (LBM) exploits both temporal and spatial parallelism, and scales the performance well with the number of FPGAs. As a result, we achieved 98.8 % of the peak performance of 73.0 GFlop/s with four FPGAs.