Putting aside, for the moment, all the Slashdot griping about whether this is or is not a productive use of human time and energy (I agree it's probably not in a macro sense, but hey, that's the world we live in), this is indeed an old idea.
I worked as a consultant on a hardware-based HFT system back in 2007 for a Silicon Valley company called Xambala. They were using reconfigurable logic-style chips of their own design specialized for text processing applications, rather than more general purpose FPGAs, though we discussing chaining their chips with FPGAs for more computationally intensive algorithms.
The edge you are going to get from doing processing in silico is quite limited. You can conceivably cut a few tens of microseconds, maybe even 100 microseconds, out of a computation - you still have to have all the other pieces of the puzzle just right. If you are doing straight news/information driven trades in situ at an exchange and can get the same timing of feed data to respond to, then you'll have a good edge (i.e. "Buy if X>0.2, Sell if X0.1, do nothing otherwise).
If you are trying to do intermarket arb (futures/ETF arb, for example) your edge is smaller, since differences in network route, networking hardware, other infrastructure are generally larger in magnitude than what you gain from cutting a few tens of microseconds out of the picture in hardware - but this edge would probably serve existing players well who already have top tier infrastructure.
For the more sophisticated, "game"-driven trading algorithms out there in equity markets, how much value doing stuff in hardware gives you is variable. There's a lot of decision logic involved in spiking orders around, changing behavior states based on other participants, and so on. A better set of algorithms running on top tier infrastructure in software will probably do better than inferior algorithms running in hardware without top tier infrastructure.
Other than Xambala, I am sure there are other players doing similar things. I've also used CUDA on NVIDIA GPUs for calculating option market prices really fast. These are just tools and other people definitely are using these tools in the right scenarios. What really matters in making money is combining the right tools with good implementation, excellent infrastructure, and testing and adaptiveness to market conditions.