Financial institutions are turning to graphics processing unit (GPU) computing for real economic and performance benefits.
Fast and accurate derivatives pricing model development and accelerated execution speeds are crucial for today's derivatives marketplace. SciComp Inc. has enhanced SciFinance, its flagship derivatives pricing software, to help quantitative developers further shorten Monte Carlo derivatives pricing model development time and create models with faster execution speeds. SciFinance now features support for NVIDIA Tesla 20-series GPUs and CUDA 3.0.
"The mathematical problems of pricing derivatives are tailor-made for GPU computing, and Monte Carlo simulations enjoy some of the fastest speed-ups on GPUs: from 50 to over 300 times faster compared to serial code," said Curt Randall, executive vice president of SciComp. "This execution speed increase makes it feasible to replace grid solutions (CPUs and interconnects) with a GPU system. GPU costs are a tiny percentage of the cost of a grid solution and offer radical reductions in both footprint and power consumption."
SciFinance takes advantage of new GPU hardware and software from NVIDIA
"Our customers can quickly take advantage of the speed increases afforded by NVIDIA's latest hardware and software enhancements," added Randall. "SciFinance automatically takes care of the CUDA programming issues. Customers need not have any CUDA or parallel computing expertise, and no hand coding is needed. All it takes is one keyword 'CUDA' added to a pricing model specification and SciFinance automatically produces optimized GPU-enabled pricing model source code."
The new NVIDIA Tesla 20-series GPUs represent a speed and feature step up from the previous generation GPUs. CUDA, a parallel computing architecture developed by NVIDIA, gives users the ability to unlock the parallel computational power of GPUs. SciFinance now supports the updated CUDA 3.0 Toolkit. At a keystroke, users can recompile their existing pricing model source code and take advantage of CUDA 3.0 for instant speed-ups in their derivatives pricing model runtimes.
New Models for Single-Tranche CDO Pricing and Commodity Futures
SciFinance has also been enhanced with new models for specific asset classes: single-tranche CDOs (STCDO) and commodity futures. The Stein stochastic recovery model for STCDO pricing has been implemented and yields results which are about 1,000 times faster than the previously used Krekel grid method. The enhanced STCDO serial code executes in about 606 seconds for an average of 5.05 millisecond per tranche-scenario (when running 10,000 scenarios on a Windows XP machine with 2.4GHz CPU). Even faster results are available with the GPU-enabled version of the STCDO code. The GPU-enabled code is about 165 times faster than the serial code, executing in about 3.67 seconds for an average of 30.6 microseconds per tranche-scenario on a single NVIDIA Tesla 20-series GPU. Thus, realistic portfolio risk computations for several thousand tranches and 50,000 scenarios could execute in O (~1hr).
A new market model for commodity futures with storage costs has been added to the SciFinance commodities example suite. Storage costs are characterized by the contango limits of the futures curve. A no-arbitrage argument is applied to determine the relation between different futures. The model is applied to price commodity derivatives of several kinds.