IP cores enable high performance FFTs in radio astronomy

July 21, 2016 // By Graham Prophet
Specialist IP vendor RFEL (Isle of Wight, UK) is supplying one of its HyperSpeed FFT IP cores to the Arizona Radio Observatory (ARO) at the University of Arizona in Tucson, Arizona where researchers are developing a high performance radio astronomy spectrometer system.

RFEL's core forms a key component to improve sideband separation in the Observatory’s heterodyne receivers. To support the receiver's wide signal bandwidth, the core utilizes parallelism to operate at a high input data rate of over 10 Gigasamples/sec. RFEL specialises in novel signal processing architectures and optimal VHDL coding that enables complex designs to utilise less FPGA resources without compromising overall system performance. In the case of ARO, the FPGA is a Xilinx Virtex 7 using Xilinx's Vivado Design Tool.


"Radio astronomy is a perfect example of how our expertise can provide solutions that cannot be found elsewhere," explained Dr. Alex Kuhrt, RFEL's CEO. "Radio telescopes generate huge amounts of data that have to be processed without loss to extract signals. This requires specialist skills to create solutions that can handle this amount of data whilst retaining the mathematical precision."


A fixed length, 8192-point, complex Fast Fourier Transform (FFT) is implemented using RFEL's HyperSpeed architecture, with an internal complex data parallelism of 32, which will be clocked at 160 MHz. The input bit format is 12-bit signed twos-complement and the input data is subject to a time-domain window function with user-programmable coefficients, allowing flexible selection of window types. The bit width will increase to 18 bits at the output of the core.


RFEL's cores can support variable bit widths for all calculation stages throughout the data path and in the ARO core. The bit widths are manipulated to ensure mathematical precision is maintained allowing the core to meet the SFDR requirements. With ARO, RFEL is optimizing the rounding schemes and bit growths within the data paths. Convergent rounding will provide the lowest Signal to Noise Ratio (SNR) but with potential structure to the noise, whilst statistical rounding will result in slightly higher SNR, but with far less structure.


Robert Freund, principal engineer at ARO, concluded, "RFEL are real experts at providing cutting edge solutions that push