I have a commercial application that requires doing an FFT convolution on a fast data stream. I currently can’t keep up the data stream in C and I would like to evaluate whether GPU acceleration is feasible. My data records are coming from a high speed ADC system and arrive as 1 million points of 32 bit integer data values. I need to do an FFT on each record, multiply the frequency transformed data by a fixed complex equalization mask and then convert back to time domain. I get 100 records per second.
I would greatly appreciate any advise for whether this is likely to be possible, how to structure the code, and which GPU would be a good candidate to purchase for the experiment. A successful test could result in several hundred GPU sales per year in our product.