What are the current computational (non-network) bottlenecks now in a quant's workflow? What computational tasks would be revolutionary with a 10-100x improvement in performance using general purpose GPUs?
closed as not a real question by Joshua Ulrich, chrisaycock, Shane, Karol J. Piczak, Graviton Feb 17 '11 at 14:12
It's difficult to tell what is being asked here. This question is ambiguous, vague, incomplete, overly broad, or rhetorical and cannot be reasonably answered in its current form. For help clarifying this question so that it can be reopened, visit the help center. If this question can be reworded to fit the rules in the help center, please edit the question.
Coming from an HPC background myself, I know too well the feeling of owning a hammer and yet having no nail. Your question is about computational bottlenecks that can be relieved with GPGPU, though I'm afraid to admit that there aren't many in finance. For realtime applications, the network is the bottleneck; for historical applications, the memory is the bottleneck. The CPU is rarely saturated in my line of work.
However, there is one particular area that does appear to be CPU bound: interpretation. Namely, the feed handler and the FIX parser both require many small amounts of data to be transformed from one representation to another. FPGA-based feed handlers are starting to become more popular; I haven't seen anything similar for FIX parsers though.
If you could show how to parse a FIX message with a GPU off the wire, then that might be interesting. FIXT 1.1 can support InfiniBand, so NVIDIA / Mellanox's GPU Direct set-up would be especially noteworthy, though not required. (There aren't many trading venues supporting FIXT right now anyway, so there's no rush there.)
If you wanted to generalize your work for all key-value pairs communicated over a network, you might be able to apply some of your findings to parsing HTTP headers in realtime. No doubt many cloud vendors would be pleased to see that.
By the way, the reason I advocated FIX parsing instead of feed handling is that most data vendors ship their own proprietary API. Good luck getting Wombat to cooperate with you until you have some results of your own to show.
In exotics options pricing, there are lots of CPU bottlenecks -- for example the calculation of Fast Fourier Transform or Monte Carlo simulation. When I price a range accrual in Libor Market Model, I don't use a lot of data (carefully optimized, everything should fit in a few MB of L2 cache), but I do a lot of calculations. This is where, I think, a GPU may be useful.