FPGA Central

Hi

I am using TI eZdsp f2812 board to do a DFT. The sampling rate is 4 MHz
so in order to get good accuracy, I need to do a 1024/2048 point DFT fo
good frequency resolution. The incoming data is continuous. I want to kno
if this can be done accurately in real time? If not, how can i make i
happen in real time provided that I have other algorithms lik
filtering/decoding running at the same time on real data.

vik wrote:
> Hi
>
> I am using TI eZdsp f2812 board to do a DFT. The sampling rate is 4 MHz,
> so in order to get good accuracy, I need to do a 1024/2048 point DFT for
> good frequency resolution. The incoming data is continuous. I want to know
> if this can be done accurately in real time? If not, how can i make it
> happen in real time provided that I have other algorithms like
> filtering/decoding running at the same time on real data.

I hope that you are planning to use an FFT and not a straight DFT!

To do a back of the envelope sanity check, you need to figure out how
how many FFTs per second are required. From your numbers, it sounds
like it could be as high as 2000. Now multiply that number by the
number of cycles per FFT, a number that is typically available in the
benchmark section of the FFT documentation (see for example TI's
DSPLIB). If that product is close to the instruction rate of the DSP
chip, you won't be able to keep up, in which case you could FFT every
other 2k segment or something like that.

Other things to watch out for include the cost of bit reversing the FFT
input, the cost of converting the FFT output to magnitude, and the cost
of memory accesses for FFT data and coefficients.

John