high performance computing on graphics processing units: hgpu.org

Since the introduction of organized trading of options for commodities and equities, computing fair prices for options has been an important problem in financial engineering. A variety of numerical methods, including Monte Carlo methods, binomial trees, and numerical solution of stochastic differential equations, are used to compute fair prices. Traders and brokerage firms constantly strive to achieve faster calculation of option prices because timely information can mean the difference between a deal struck or missed, which translates to substantial profit or loss. Hence, the latency to compute a fair option price plays an important role in short-term trading and arbitrage. Financial firms constantly seek faster, more accurate option pricing methods in order to keep up with or ahead of their competitors. Approaches to improve latency include both more efficient trading strategies or pricing algorithms and use of specialized, high-performance computer architectures, such as FPGAs, many-core CPUs, and GPUs. An exemplary low-latency implementation on FPGA of European Options pricing via Monte-Carlo simulation was described in [1], where 15x speedup over an existing server, as well as outperforming a GPU and Cell implementation was reported. Monte-Carlo methods are usually exhaustive, time and compute intensive and take into consideration various sources of uncertainties corresponding to real market conditions. On the other hand, lattice(esp. binomial tree) methods are faster and consider relatively few possibilities, uncertainties and is a reasonable approximation to a variety of standard market conditions.