In order to quickly develop new structured products and vastly speed up Monte Carlo-based derivatives for its Software-as-a-Service (SaaS) customers, Reval, the global leader in derivative risk management and hedge accounting solutions, chose SciFinance(R) from SciComp Inc.
SciComp, Inc., a leader in automated code generation software, is exhibiting at the annual ICBI Global Derivatives & Risk Management Conference in Paris this week. Attendees can learn more at SciComp's talk on May 20th at 12:10 pm, "Automatic GPU Computing For Derivative Pricing Models."
SciFinance leverages the power of high-end NVIDIA GPUs (Graphical Processing Units) to accelerate derivative pricing performance. It automatically generates GPU-enabled pricing model source code for any Monte Carlo-based model.
"We were looking for a cost-effective and easy-to-deploy solution to improve the pricing of complex derivative instruments using PDEs or Monte Carlo simulation in our SaaS product. We found it with SciFinance and GPU-enabled models, without having to become experts in parallel coding or CUDA," said Ernest Bonsell, Senior Vice President, Product Engineering at Reval.
Monte Carlo simulation requires running between 20,000 to 50,000 trials and uses a great deal of CPU time. Implementing a Monte Carlo framework in a SaaS environment adds extra stress to CPU usage at month-end closing and at times of high system usage. With over 400 clients and 1,325 users, processing speed is an important consideration for Reval.
"Reval benefited immediately from SciFinance's ability to generate GPU-enabled code," said Curt Randall, executive vice president of SciComp. "They can now quickly output code that delivers an immediate 50-300X increase in execution speed."
SciComp has integrated optimal CUDA coding paradigms into SciFinance to take advantage of the parallel processing power in GPUs. For Monte Carlo computations, instead of the CPU computing one path after another, 50,000 times in serial fashion, the CPU can send the GPU requests for each of 100 cores to compute 500 paths and then collate results.
"While SciComp has invested significant effort in producing highly-efficient parallel code for execution on NVIDIA GPUs, from Reval's point of view it couldn't be simpler. Financial engineers simply add the keyword "CUDA" to an existing serial specification and regenerate the parallel code in minutes," added Randall.