high performance computing on graphics processing units: hgpu.org

In this paper, we propose, develop and implement a tool that increases the computational speed of exotic derivatives pricing at a fraction of the cost of traditional methods. Our paper focuses on investigating the computing efficiencies of GPU systems. We utilize the GPU’s natural parallelization capabilities to price financial instruments. We outline our implementation, solutions to practical complications arising during implementation and how much faster GPU systems are. Each step that we explore has a significant impact on the efficiency and performance of GPU pricing. Rather than speaking in theoretical, abstract terms, we detail each step in an attempt to give the reader a clear sense of what’s going on. Efficiency is one of the pillars of financial calculations. With the volume of risk calculations mandated by prudent risk management practices, even moderate improvements in calculation efficiency can translate into material changes in trading limits or savings in regulatory capital. This can make the difference between a growing, successful trading operation or an also-ran. Unfortunately, a decent algorithm written in VBA cannot calculate option prices at the same speed as a farm of computers, particularly if we must price the trade in less than 150 milliseconds using 10 million simulation paths. Fast forward from one trade to a book of several hundred thousand trades, many of which are exotic products. Not only is it necessary to price each trade, but we must do so in each of thousands of different market scenarios in order to calculate even basic risk measures such as Greeks and Value-at-Risk (VaR). At the end of the paper, we discuss how GPUs are currently used in the industry and their various advantages, including cost, time, accuracy and calculation frequency. In addition, we discuss the implementation challenges of GPU systems and the attention to detail that is required for memory allocation.