Long FFT; DSP or FPGA?

[jeremy]

Hi all,

I'm trying to design a device that will calculate a long FFT (~100k point sampled at 1M) and I'd like to calculate it at 1-2Hz on an embedded system that can run off a lipo stack. I am currently at a loss as to whether to go for an FPGA or a DSP to do the calculation, has anyone done this before and can offer some advice?

[jeroen74]

I think a DSP is easier, but that depends a bit on the design of the rest of the system. There's more to the FPGA then just the stuff to do the FFT, think interface to memory, interfacing to the other parts of the system etc.

[EEVblog]

Depends entirely on how you implement the DSP algorithm. If you don't do (or don't know how to do) parallel tricks on the FPGA, then the DSP often will be quicker and easier.
A lot of people make the mistake of just using a soft core processor in the FPGA to run the same algorithm as a micro/DSP and wonder why it's slower 

Try:
http://www.altera.com/literature/wp/wp_dsp_comp.pdf
http://www.hunteng.co.uk/info/fpga-or-dsp.htm
http://www.ieee.li/pdf/viewgraphs/mapping_dsp_algorithms_into_fpgas.pdf

[AndyC_772]

How will you get the data into the device that's doing the processing?

[jeremy]

The samples are going to come in via the FPGA, so that's why I was considering an FPGA. Trouble is, I don't have a lot of experience with embedded DSP beyond simple FIR filters, so while I get the theory, I am lacking in knowledge of the "tricks" that Dave mentioned for FFTs. I guessed (correctly I suppose?) that DSPs would be easier and there will probably have to be a processor of some sort on the board  anyway.

I wish these hybrid ARM/FPGA SoCs weren't in such horrible BGA packages  They seem perfect.

I should clarify: I'm interested in the power spectral density aka real FFT, not all that interested in the phase.

[cosmos]

How often would you like the FFT to run?
If the DSP can handle the rate I would not bother with FFT in FPGA.

[jeremy]

Basically just 1-2Hz, acquisition takes 100ms anyway.

Also, I have had trouble in the past with TI not supporting their chips very well (bought a dev board from them, needed to go back to XP to use it!). Can anyone recommend another chip with more open source and/or well supported tools?

[EEVblog]

Basically just 1-2Hz, acquisition takes 100ms anyway.
Also, I have had trouble in the past with TI not supporting their chips very well (bought a dev board from them, needed to go back to XP to use it!). Can anyone recommend another chip with more open source and/or well supported tools?

I would just get a DSP demo board and give it a try using dummy data maybe and see how it goes. If it's good enough, then don't bother with the idiosyncrasies of fpgas.

[jeroen74]

Maybe you should take a look at the Blackfin series from Analog Devices. I have no experience with them, but a quick view at Farnell shows they are available in reasonable packages at reasonable prices too.