Let's compare the number of operations needed to perform the convolution of
2 length sequences:
N | FFT | Direct Convolution |
4 | 176 | 16 |
32 | 2560 | 1024 |
64 | 5888 | 4096 |
128 | 13,312 | 16,384 |
256 | 29,696 | 65,536 |
2048 | 311,296 | 4,194,304 |
In this example (from Strum and Kirk), the FFT (software) beats direct time-domain convolution at length 128 and higher