Slashdot Mirror
JPMorgan Rolls Out FPGA Supercomputer
An anonymous reader writes "As heterogeneous computing starts to take off, JP Morgan have revealed they are using an FPGA based supercomputer to process risk on their credit portfolio. 'Prior to the implementation, JP Morgan would take eight hours to do a complete risk run, and an hour to run a present value, on its entire book. If anything went wrong with the analysis, there was no time to re-run it. It has now reduced that to about 238 seconds, with an FPGA time of 12 seconds.' Also mentioned is a Stanford talk given in May."
Do not click. Some sort of phony coprophile spam.
It seems unlikely that they're doing anything that would prevent them from doing it significantly faster with less power draw on a GPU, and with lower initial hardware costs to boot.
No, FPGAs use significantly less power and provide greater performance than GPUs. The initial capital cost is higher though. Here's an article that gives a bit more detail: http://www.xilinx.com/publications/archives/xcell/issue74/FPGAs-speed-computation-complex-credit-derivatives.pdf
That's not how it works. The Maxeler FPGA compiler is actually a Java library for describing dataflow machines. They call it "stream computing" - basically just very deep pipelines. You describe your algorithm in the dataflow/streaming style, using the Maxeler Java library, then when you run your program the output is an HDL design to load onto the FPGAs. There's no automatic conversion from a serial program (in C++ or anything else) to a dataflow/streaming program. The developer reimplements the algorithm manually.
Well, why don't you? The beauty if VHDL is that it's really, really basic. The rest is up to you to implement. Come join us in ##vhdl at irc.freenode.net or try fpga4fun.com. $100 should get you an FPGA training board which you can use to learn.