Using Graphics Power For Financial Engineering
September 8, 2009
Computers powered in part by graphical processing units are increasingly making inroads into automated trading applications. The end goal of chip makers such as Nvidia is to migrate supercomputers from the data center to the desktop.
Nvidia, whose chips earned their stripes inside personal computers used by ardent video game players, says it's done that already with its Tesla line of graphics processors.
In May, the first Dell Precision R5400, T5500 and T7500 personal computers shipped with Tesla GPUs available. This combination, the firms said, turn the systems into "personal supercomputers" which could be used for financial applications. The units can cost up to $10,000 a piece and each Tesla card can execute near a trillion calculations a second.
Applications already using Nvidia's CUDA programming language to crunch values include SciComp's SciFinance derivatives pricing engine, New York-based Hanweck Associates Volera real-time options valuation engine, Houston-based Aqumin's 3D market data visualization module, a risk analysis system in St. Louis-based Exegy's ticker plant and a trinomial options pricing engine at Boulogne-Billancourt, France-based proprietary trading firm, Arbitragis Trading.
Oxford University's Skynet, for instance, a Nvidia-based cluster of four of Tesla S1070 units is "something not much bigger than your desk," says Robert Meyer, CEO at Numerical Algorithms Group (NAG). According to SciComp, one PC equipped with several Nvidia GPU cards can replace multiple racks of blades. A Tesla C1060 board, for instance, puts 240 graphic processing units into a space 4.4 inches wide by 10.5 inches long.
The Texas Advanced Computing Center at the University of Texas is announcing this week a new system that blends the power of graphics processors and central processors, said TACC's Melyssa Fratkin.
The challenge with GPU-based computing, Fratkin said, is getting coders to learn and firms to adopt a new programming language. In Nvidia's case, that's CUDA, or the Compute Unified Device Architecture, which is a variant of the "C" programming language.
"My director is teaching Fortran in C on campus today as we speak, because the computer science departments don't teach that stuff anymore, but you can't program supercomputers without it," Fratkin says. "He's got 30 people in that class: It's full."
Most GPU-based systems now work in conjunction with the traditional power sources for supercomputing, central processing units (CPUs), as well as with field-programmable gate arrays (FPGAs), which are programmable or customize-able integrated circuits often favored by financial ticker plants. Exegy, for example, uses all three methods-CPU, GPU and FPGA-simultaneously to perform Monte Carlo analyses. Doing so boosted the speed of the simulation by 310 times versus prior architectures, according to Exegy.
Merle Giles, director of the private sector program at the National Center for Supercomputing Applications (NCSA) at the University of Illinois at Urbana-Champaign, says NCSA has the largest deployment of a CPU- and GPU-powered supercomputer in the country. NCSA's Lincoln supercomputer has 96 Tesla S1070 accelerator units; each Dell server is connected to two Tesla processors.
Not everyone is convinced GPU-based processing has proven its long-term value down to the end-user. "Banks want commodity hardware," says Neil Bartlett, chief technology officer at Toronto-based Algorithmics, a financial risk management vendor whose clients include Bank of America and Deutsche Bank. "You have to have banks that believe that Nvidia's going to be around forever, whereas it's easier to look at a chip manufactured by Intel because they've been around the space."
