By Chris Kentouris
BNP Paribas, moving to bolster its computational power, has implemented a new technology platform designed to not only accelerate calculation times for complex equity derivatives but also slash energy consumption.
Last week, BNP Paribas completed the transfer of risk management calculations for about 5 percent of its global equity derivatives portfolio from Intel Corp. central processing units (CPUs) to Nvidia technology that allows for hardware acceleration based on graphics processing units (GPUs).
To speed up the implementation--conversion took about six months--the French investment bank opted to limit the scope of the project. "We were using about 10,000 cores of Intel processors on Ada code and decided to move 5 percent of the CPU cores to eight Nvidia GPUs on two Tesla S1070 series computers," explained Stephane Tyc, global head of quantitative research for equity and commodity derivatives.
The Nvidia GPUs are performing value-at-risk (VAR) calculations using Monte Carlo simulations, a complicated, though common, way to determine a portfolio's maximum loss over a certain period of time. About 20,000 lines of code are involved-or 2 percent of the total code used by the equity derivatives unit.
Adding grids and servers to achieve high-performance computing can be costly and drive up energy usage. Through GPU acceleration, a relatively new approach, firms have been able to both rapidly process market data and analytical calculations and reduce rack space at data centers. GPUs from vendors like Nvidia and Advanced Micro Devices break down large problems into smaller ones that can be processed simultaneously.
"GPU acceleration is an emerging and largely unproven technology," noted Adam Honoré, senior analyst with Boston-based Aite Group. Still, many large financial firms are investing in it.
Tyc, who was convinced of the technology's merits by Florent Duquet, a consultant with a PhD in computer graphics and experience with GPUs, said he received approval for the initiative in March 2008 from Olivier Osty, deputy head of global equity and commodity derivatives. "We did a proof of concept quickly and found [Nvidia] to be the best available solution out there on the GPU technology front," said Tyc.
Nvidia's processors, which initially supported three-dimensional video game graphics, are now used in the oil and gas, medical, manufacturing and life science markets. The Santa Clara, Calif.-based company entered financial services two years ago with Tesla workstations that accommodate the need for high-speed pricing and VAR calculations. John Milner, Nvidia's director of business development, said the average client uses as many as 32 of the Tesla GPUs in the initial stages of a project.
"Tier-one investment firms have maximized their use of data centers and see the advantages of reducing their overall power consumption and real estate," said Milner.
Compared to a dual- or quad-core CPU, the newest Tesla GPUs have 240 cores. According to Tyc, they increase by 100 times the number of calculations BNP Paribas completes per watt of energy. That, he said, translates into 190-times less overall energy consumption.
Dominique Le Campion, head of computational architecture in the BNP unit, added that by providing traders with the calculations intraday, "rather than each morning after an overnight batch process is run, trade executions can be done on a faster basis." And, he noted, the GPUs offer more accurate results by allowing the inputs used in the Monte Carlo simulations to be adjusted.
GPU vs. FPGA
On Wall Street, GPU technology has been gaining on the more established field-programmabl gate array (FPGA) hardware. While both apply parallel architectures, FPGAs primarily have been used to reduce data latency in trading feeds. GPUs are more easily programmed, according to Aite's Honoré, which means they can be applied more quickly to changing requirements than the fixed-code FPGAs.
"There aren't enough quantitative analysts specializing in FPGAs' proprietary coding," said Jeff Wells, VP of product management for Exegy. The St. Louis-based ticker plant specialist uses FPGAs to cut down on latency in processing multiple data feeds.
Exegy ran a test in November using CPUs, FPGAs and GPUs to perform Monte Carlo calculations on a portfolio of 1,024 equities. According to Exegy, "a calculation that would normally take 15 minutes on a multicore CPU now only takes 12 seconds with all three technologies." The study did not directly compare FPGAs and GPUs.
Nvidia has also been tapped by third-party software developers such as derivative pricing specialist SciComp and Hanweck Associates, a New York-based quantitative financial consulting firm, to develop applications for fixed-income and derivatives trading desks. In 2007, Hanweck incorporated the GPU technology for low-latency options analytics.
For BNP Paribas, its conservative use of GPUs is less a reflection of their capabilities than of its desire to capitalize on any new technologies that pop up over the next few years. "Maybe the supercomputing environment will have changed by then and our work will become obsolete," said Tyc. "We want to nibble in different directions and so far haven't identified the next big winner."
The bank's fixed-income derivatives unit is currently considering switching from Intel CPUs to Nvidia GPUs.
This article can also be found at SecuritiesIndustry.com.
Register or login for access to this item and much more
All Information Management content is archived after seven days.
Community members receive:
- All recent and archived articles
- Conference offers and updates
- A full menu of enewsletter options
- Web seminars, white papers, ebooks
Already have an account? Log In
Don't have an account? Register for Free Unlimited Access