Monte-Carlo simulations are among the most common numerical methods in computational finance. They are used when closed form solutions or other numerical methods are not practical or do not exist. Complex financial problems are often evaluated numerically using Monte-Carlo methods. For example, Monte-Carlo derivative pricing determines the current value based on a large number of random experiments and some form of statistical analysis to obtain the results.
This video demonstrates the benefits of using the Xcelerit platform to efficiently run Monte-Carlo simulations using multi-core CPUs and graphics processing units (GPUs). The generic approach for Monte-Carlo simulations using the Xcelerit platform is presented and performance figures are given for the specific examples of pricing European options and pricing a portfolio of LIBOR swaptions.
The performance of the Xcelerit platform implementation on GPUs as well as an equivalent sequential implementation on a single CPU core has been measured for the execution of the full application. The test system for benchmarks is configured with two Intel Xeon E5620 CPUs (each with 4 hyperthreaded cores, i.e., 16 hardware threads in total), 24GB of RAM, Red- Hat Enterprise Linux 5.4, and two Nvidia Tesla M2050 GPUs (ECC off).
The speedups using both GPUs compared to the sequential reference are between 80 and 150x, depending on the number of options and the number of paths used (single precision). This corresponds to around 1 billion paths/second. For large numbers of simulation paths (>100K), the result matches the solution of the closed-form Black-Scholes formula.
Full white paper in PDF is available at http://www.xcelerit.com/monte-carlo-methods/ (Registration required!)
Note from GPU Science: One should take claimed 150x and even 320x acceleration with a grain of salt. Correct comparison of one GPU vs one CPU (4 cores) or even one GPU vs Full CPU box (8 cores) would result in much more realistic speedups on the order of 10x – 30x (depending on the number of options).
[Submitted by Hicham Lahlou, Xcelerit]