(snip, I wrote)
As far as I know, currently if something can be done either with an FPGA or a small microcontroller, or even DSP processor, the choice is often for the processor. With the price of smaller FPGAs coming down, that might change, but also there are more people who know how to program such processors than HDL design.
Given that, I would expect some overlap between FPGA/HDL applications and processor/software applications, depending on the speed required at the time. Some applications might initially be developed in software, debugged and tested, before moving to an FPGA/HDL implementation.
Maybe the question isn't whether scaled fixed point is useful in HDL, but why isn't it useful, enough that people ask for it in an HLL, in software! Why no scaled fixed point datatype in C or Java?
(snip)
(snip)
As I said above, if for no other reason, then to test and debug the applications that will later be implemented in scaled fixed point VHDL.
Well, there are two applications that D. Knuth believes should always be done in fixed point: finance and typesetting. As we know, it is also often used in DSP, but Knuth likely didn't (doesn't) do much DSP work. (It was some years ago he said that.)
(snip)
Maybe so. In the past, the cost and size kept people away from using FPGAs when conventional processors would do. With the lower prices of smaller (but not so small) FPGAs that might change.
(snip)
Yes. If I determine the shifts and rounds, then I don't need VHDL to do it for me. I can do it using integer arithmetic in C, (easier if I can get all the product bits from multiply), and in HDL.
(snip)
Well, the desire to run almost any computational problem faster certainly isn't unique to me. But yes, I know the problems that I work on better than those that I don't. But if you can't process data faster than a medium sized conventional processor, there won't be much demand, unless the cost (including amortized development) is less.
True, but for floating point number crunching, in the teraflop or petaflop scale, it is scientific programming. With the size and speed of floating point in an FPGA, one could instead use really wide fixed point. Given the choice between 32 bit floating point and 64 bit or more fixed point for DSP applications, which one is more useful?
In general, fixed point is a better choice for quantities with an absolute (not size dependent) uncertainty, floating point for quantities with a relative uncertainty.
-- glen