"Again, if you are doing it right, you never use floating numbers in finance. Every calculation is done with integers, and you keep track of the number of decimals separately. No rounding will occur. No floating numbers are needed. As I said, I work in a large finance company."
Hmm, well I'm not sure that you've wholly understood what that part of POWER is. As you say, floating point is no good. I gather that the *decimal* accelerator (not the FPU) in POWER is doing the sort of arbitary precision sums you've outlined. Hardware acceleration of that is obviously going to bring benefits, or so IBM would have you believe. And it's difficult to disagree with the evidence of their sales figures.
"Those are trivial calculations, done in COBOL on Mainframes. Not very sexy."
Not sexy, but clearly very profitable. Profits don't have to be earnt in a sexy way, they just have to be big! I'd say that on an absolute scale skyscrapers, suits and MBAs are not really significantly sexier...
Yes, I'm familiar with the technology that the high speed trading world uses, and I'd certainly agree with you on the inappropriateness of mainframes in that role! But generally I think that Solaris/Linux on even top end server and network hardware is behind the curve when it comes to low latency, largely because they're stuck with stodgy sluggardly interconnects like Ethernet, Myrinet and Infiniband.
The high performance embedded signal processing world has been much more focused on low latencies than the mainstream server world. The 'unconventional' interconnects found in that domain (VXS's and OpenVPX's sRIO, and external interconnects like sFPDP) are all about low latency. That's because it's a key driver in the sorts of applications (radar, etc) implemented using such hardware. If you've not done so already, it's worth a bit of investigation.
Anyway, hurry up and earn your profits. There's a good chance that the whole high speed trading thing will get banned soon, especially if it gets fingered for causing a major market wobbly. I'm pretty sure that no one in the finance industry could say whether or not it meets the Nyquist stability criterion, but those of us who know what that means generally think that it doesn't and don't believe that anyone's checked to see either.
Even if it doesn't fall over it's doomed stagnate, eventually. As soon as you've all bought premises as close as possible to the stock exchange and have all chosen the optimum hardware and algorithms for the job, you'll all be as good at it as each other at it and there won't be any technological advantage left to exploit. Anyone started checking to see if they're plugged in to port number 1 on the Exchange's Ethernet switch?
On the otherhand if stagnation spurs the finance industry into developing even lower latency kit (the mainstream IT industry won't, they care merely about throughput) then I would be quite grateful.