A FPGA can make a DSP look like a sick puppy, that is one reason I
recently bought a Spartan 3 1000K board, 24 multipliers in hardware
built in. I looked for examples and found a 4096 point FFT using
floating point in 7.6 microseconds, is that fast enough? We would not
need it anywhere that fast so we could cut back on the hardware usage.
FPGA-based DSP is useful for certain applications where lots of speed is
needed for a straightforward algorithm, like the FFT you cite.
Real-time video filters, etc., are fertile ground for FPGA-based DSP.
But it is very difficult to fit into the same FPGA -- running at the
same time -- a denoiser, adjustable filters for SSB and CW, a Tx
equalizer, impulse noise blanker, SAM tracking demodulator...
A general-purpose DSP is useful for applications that require a bit less
raw horsepower but great flexibility in the algorithms being performed.
A good example is the SDR-1000. The DSP being done in the code running
on the PC would likely be be a nightmare in an FPGA, unless you created
a "soft" programmable DSP inside the FPGA, but that's cheating...
I create designs with DSPs and FPGAs, and use both extensively. Each
has its place -- I want to be sure there is a place in the "not Xylo no
acronyms sorta universal widget box" being discussed, even if that place
is never occupied :-)