I have an application where I have to run a 2048 or a 4096 points FFT every 30 seconds. I am wondering whether I can accomplish this with a regular microcontroller since my timing requirements are not that stringent. It is not greatly important whether the FFT is integer or floating point, the values are 16 bit A/D values. I was considering just using a plain old TI-MSP 430 with 10K or RAM. Does anyone have a better suggestion as far as ease of implementation, low power and price?
Since that is at the top-end of MSP430, I'd also look at the bottom-end of the many ARM Flash uC recently appearing. Philips LPC2xxx family would be a good place to start, but you could define more peripherals, and that would focus your choice more. AnalogDevices, Atmel, Intel, Motorola, Philips, ST are all busy announcing/ramping ARM single chip uC in various market segments.
I suggest getting a portable C implementation and trying it on the MSP430. I haven't tried it, but if your processor has the built-in multiply and you use a sine lookup table, I suspect it will do the transform in a few seconds, under you 30 second allotment.
Your first problem with doing an FFT is data space. A 4K transform is going to need at least 16 Kbytes for even a single-precision float form. This won't fit well in 10 Kbytes of RAM.
using the LPC2105 or LPC2106 from Philips with 32 / 64k RAM would definitely be an easier implementation than trying it with a 10k MSP
430. However, if stand-by current is an issue (battery driven?) you might want to try the MSP430 as I consider it best in class for stand-by. Looking at Mips/Watt I would think the ARM devices can actually easily compete.
I guess it depends on how you define "combo" - Those listed above and the lower-end TI, Motorola (and I presume AD) DSP's are basically DSP cores with embedded peripherals added on-chip, and perhaps also advertised with extra "embedded instructions" for fast bit twiddling. There seem to be many of them around, so perhaps these should have a different name to distinguish them from plain-vanilla high-end DSP's. Maybe it should be called a "Digital Signal Microcontroller." or DSM A DSM has a MAC and instructions to use it, as well as the features of microcontrollers (data RAM, program ROM, I/O ports) so that it can be made into an operational computer (with "DSP capabilities") with only a small addition of external passive components. The distinction between microprocessor and microcontroller was made many years (decades?) ago (who did it? Was it Motorola with the
68HC11, or maybe Intel's 8048?). DSM's have grown in popularity without being distinguished from the original DSP's.
Zilog had (and may still have) a "true" combo of both a microcontroller and a separate DSP on one piece of silicon, with various ports so they can communicate with and (presumably) interrupt each other. I haven't heard of any other such device.
Good point, Motorola even tries the strange label of 'hybrid', which does not gel well in a industry that already knows what hybrid means. I give dsPIC a higher name selection score.
There are quite a few at the top-end, where DSP+ARM is now almost the default for certain cell phone engines, tho these are not single-die system solutions in the way a microcontroller is.
My specs dictate that I have a 14-16bit ADC converter, this can be internal to the processor or external. The requirement is the ability to read at least 256 samples per second, ideally 512 samples per second. We would repeat the reading for approximately 8 seconds and perform some FFTs on the data. The philips LPC2106 seems like a good option however it lacks an internal A/D. Furthermore, are their A/D converters capable of acquiring that many points at that resolution that quickly?(internal or external).
There are plenty of 16-bit audio D/A converters that sample 48,000 times a second. I don't know off hand of microcontrollers with internal ADCs wider than 12 bits, but I don't know much. I don't know what you intend to do with the data you collect, but be aware that many converter/- anti-alias filter combinations impose delays of many sample periods. I would call a sample rate below 1 KHz relatively slow.
What do you mean by repeating the reading? You could mean acquiring a single string of 8 x 512 = 4096 contiguous samples and do one FFT, or else acquire 8 strings of 256 samples each, perhaps with caps between them, and doing 8 shorter FFTs, presumably averaging the results. Since the FFT gives cosine and sine (or real and imaginary, as you please), and averaging requires converting that to magnitude, the first way is likely to need less computation.
Jerry
--
Engineering is the art of making what you want from things you can get.
¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯
This may sound silly bit I ask if you need a full FFT. Do you want to find the magnitudes of ALL frequencies in your range or just a few frequencies of interest. If it's the latter then you don't need FFT just synchrronus demodulation and that's alot simpler.
I can point you int that direction if it would help.
ElectronDepot website is not affiliated with any of the manufacturers or service providers discussed here.
All logos and trade names are the property of their respective owners.