Rough performance estimate:
for a 1/10 transition width filter (which is sufficient to keep 20 of 25 MHz Nyquist), you
need an expertly guesstimated [1] 24-ish taps, so go for 25 taps: that happens to be
exactly the minimum filter length that we can use in an 25-interpolating rational
resampler and still do the full polyphase decomposition trick to run the filter at the
least rate of the system.
For "1 tap per branch", I have a rough 60 MS/s for *my PC* in mind. So, My guess is that
it will work for 46->50 MS/s, not for 92->100 MS/s, on *my PC*.
Problem: this is an inherently badly multi-threadable workload, unless I'm overlooking
something. So, having 48 cores isn't better than having 2.
Just curious, why do you think the FFT interpolation won't work?
So long as the block sizes are appropriately picked (how about we stick with 4600 input size, and 5000 output size), the frequency components should all line up and be phase aligned from block to block. There might be a little scaling factor that the resulting samples will be off due to the difference in FFT length, but otherwise I think it should work.
Moreover, the FFT work can be done in parallel and utilize all the cores.
I did a quick little test and I think it should work. Yes?
Brian