Automatic generation of high-throughput systolic tree-based solvers for modern FPGAs

Abstract

Tree-based models are a class of numerical methods widely used in financial option pricing, which have a computational complexity that is quadratic with respect to the solution accuracy. Previous research has employed reconfigurable computing with small degrees of parallelism to provide faster hardware solutions compared with general-purpose processing software designs. However, due to the nature of their vector hardware architectures, they cannot scale their compute resources efficiently, leaving them with pricing latency figures which are quadratic with respect to the problem size, and hence to the solution accuracy. Also, their solutions are not productive as they require hardware engineering effort, and can only solve one type of tree problems, known as the standard American option. This thesis presents a novel methodology in the form of a high-level design framework which can capture any common tree-based problem, and automatically generates high-throughput field-programmable gate array (FPGA) solvers based on proposed scalable hardware architectures. The thesis has made three main contributions. First, systolic architectures were proposed for solving binomial and trinomial trees, which due to their custom systolic data-movement mechanisms, can scale their compute resources efficiently to provide linear latency scaling for medium-size trees and improved quadratic latency scaling for large trees. Using the proposed systolic architectures, throughput speed-ups of up to 5.6X and 12X were achieved for modern FPGAs, compared to previous vector designs, for medium and large trees, respectively. Second, a productive high-level design framework was proposed, that can capture any common binomial and trinomial tree problem, and a methodology was suggested to generate high-throughput systolic solvers with custom data precision, where the methodology requires no hardware design effort from the end user. Third, a fully-automated tool-chain methodology was proposed that, compared to previous tree-based solvers, improves user productivity by removing the manual engineering effort of applying the design framework to option pricing problems. Using the productive design framework, high-throughput systolic FPGA solvers have been automatically generated from simple end-user C descriptions for several tree problems, such as American, Bermudan, and barrier options.