The block diagram of the FFT subsystem is shown in Fig. 25.4. The basic unit of the FFT subsystem takes two data streams (either A,B or C,D in Fig. 25.3) from the Delay-DPC. In the first stage, a weighting function can be applied to the 4 bit time series. The weighting function is software selectable, and can be chosen to be one of the standward ``window functions'' discussed in Chapter 8. This is followed by a number controlled oscillator (NCO), which does the fringe stopping (see Chapter 9). The two fringe stopped time series are passed through two sets of FFT engines, realized using VLBA ASICs, to perform Fourier transforms. Phase gradients are then applied to the spectrum of the signal to correct for delays smaller than the sampling interval (FSTC).

Each FFT engine can perform a Fourier transforms of maximal length 512 points. This length is software selectable to be 256, 128 or 16 points; it is even possible to bypass the FFT operation altogether. A 512 point FFT gives 256 channels, however in the next stage of the correlator (MAC) there are only enough multipliers for 128 channels per sideband per polarization. In the standard mode of operation, two adjacent FFT channels are hence averaged together in the MAC. A single MAC also acquires data from two FFT engines in a time multiplexed fashion. The data is multiplexed as shown in Fig. 25.4, where and are the spectral channels from the two FFT engines.