Sam Brown of the University of Utah just came out with
some slides on his initial forays into hardware acceleration. He contrasts the use of Cell Processors,
FPGAs, and
Multi-core processors. He was trying to speed up 2-Way Wave Equation Modeling, which is not a financial application, but the lessons that he learned can be applied to those of use who are doing (or exploring)
HPC in Wall Street.
The big effort is in unrolling loops and
parallelizing your code. Using the
POSIX PThread API and
multi-core
cpus is the easiest thing to do. With 4 threads (2 dual-core 2.2
Ghz Opterons, 4 GB RAM), he got a 11x improvement on one set of benchmarks, and a 2x improvement on the big application.
The
FPGA system used the
Altera Stratix II
FPGA. He got about a 14x
improvement on the initial set of benchmarks, but the development effort was a lot harder. It took 12 hours to get the design onto the
FPGA chip.
The Cell systems were a Sony
Playstation 3 (!!!) and a IBM Cell Blade with 2 Cell Processors with 8
SPEs each (a Cell processor has 8
SPEs, which function as coprocessors.) The Cell results were generally disappointing, with the author wondering if he programmed the Cell correctly. One thing that I have heard about the Cell is that it is very difficult to program. Perhaps companies like
PeakStream can help you with this effort.
(For instant gratification, the money shots are slides 28, 29, and 30)
What can hardware acceleration do for you? Faster pricing of derivatives. Faster generation of risk values, especially for exotics. Better overnight stressing.
The problem is in acquiring the
skill set to
parallelize algorithms, deal with the development environments (especially debugging), learning new variants of C, and getting buy-in from the business.
©2007 Marc Adler - All Rights Reserved
