FROM EDN EUROPE: Channelise wideband signals in real time without DSPs

RF Engines has developed the PFT (Pipelined Frequency Transform) method of channelising wideband signals (Picture). The method is as much as 20 times more efficient than using an array of digital-frequency converters or an FFT pipeline with massively parallel processing. The basic function is that of a spectrum analyser, subdividing an incoming signal as wide as 100 MHz into a number of discrete channels, or bins. You might need such a function in cellular-phone base stations, radar or surveillance equipment, or in instrumentation. The PFT is a hardware-based solution that you buy as IP for implementation on SOC ASICs or FPGAs.

You might conventionally approach such a task with a "tree" structure of downconverters. Given an input signal I/Q form at some given sampling rate, you would divide its bandwidth in two, then divide each of those in two again, and so on. At each stage, you would yield twice the number of output channels, each of half the bandwidth of those in the previous stage, and with the data at half the sampling rate. However, as your required number of channels rises, the required number of complex frequency-converter blocks rises geometrically. The PFT hardware structure also uses successive stages of frequency conversion, but each stage requires only two PFT downconverter modules, one in each of the I and Q paths. The sampling rate stays the same throughout, and, after each stage in the process, there is information in the I/Q datapath on the content of twice as many subdivided channels as there was in the stage before. This data is now interleaved in the bit stream. The original information on the wideband signal is still present, but, instead of a single fast sample stream covering the whole band, the processing converts it into a set of interleaved bit streams, each representing the content of the divided channels.

The frequency-conversion block in this architecture uses FIR filters. Adding delays in the FIR-filter structure performs interleaving. Thus, the bandpass characteristic for each channel can be excellent. Although you could also perform such a channelisation with a pipelined FFT, achieving the same filter performance would be difficult.

To achieve the 20-fold efficiency increase, RF Engines tested the approach with a 100-MHz bandwidth at 8-bit sampling resolution, extracting 1024 channels with sharp filter characteristics. You could see filter-stopband rejection of 75 dB with gain across each bin flat to within 0.2 dB. This approach would fit in four Xilinx Virtex FPGAs; moreover, according to RF Engines' chief technical officer John Lillington, you can run this 1024-point transform 20 times faster than an FFT implemented on a DSP. The further you subdivide the band, the greater the advantage over a digital frequency-converter structure; a 16,000-point transform is practical and would need 14 PFT stages, as opposed to 16,384 individual frequency converters. Because of the pipelined structure, there is no lost data. You may need to employ postprocessing, but, depending on the real-time-processing power at your disposal, you can track signals moving at hundreds of gigahertz per second and identify hopping and other fleeting signals that appear only in tight channels for microseconds.

RF Engines delivers the PFT through a four-stage IP-licensing model; it takes the form of a relationally placed macro operating as fast as 80 MHz on VirtexE FPGAs with full supporting documentation.

RF Engines, +44 1983 550330, www.rfel.com.