Available Technologies

Background

This technology was developed for radio observation spectrometers in the field of astrophysics, enabling wideband, high-resolution real-time FFT processing with lower memory usage than conventional methods.
In radio astronomy, spectrometers capable of high-resolution and highly efficient spectral analysis of radio signals are required. Specifically, FFT technology capable of simultaneously processing several GHz of bandwidth with frequency resolutions on the order of several tens of kHz is demanded. This requires processing extremely large-scale input data corresponding to FFT sizes on the order of [1 x 105]-points. In conventional FFT implementations on GPUs and FPGAs, twiddle-factor memory requirements increase rapidly as bandwidth and frequency resolution increase, leading to challenges such as insufficient FPGA resources and increased power consumption..

Description and Advantages

The inventors focused on the exponential properties of twiddle factors and devised a rotation-processing circuit that can be implemented with minimal memory and arithmetic resources. Furthermore, they developed a novel FFT architecture combining this circuit with a proprietary structure for efficient utilization of twiddle-factor tables, and implemented it on an FPGA.
Compared with conventional approaches, the proposed technology significantly reduces memory usage while minimizing the increase in multiplication operations, thereby enabling low-memory, wideband, high-resolution real-time FFT processing.

⮚Reduced Memory Usage Contributing to Lower Power Consumption and Smaller Circuit Scale
•Twiddle-factor memory reduced to approximately 1/8-1/16of conventional implementations
•Overall FFT circuit memory reduced by approximately 20-40%while limiting the increase in multiplication operations to approximately 10% compared with conventional methods

⮚Real-Time Processing of Wideband and High-Resolution FFT

Demonstrated specifications of the "dSpec" spectrometer incorporating the proposed FFT algorithm:
•FFT size: N=217～18
•12-bit input width
•4 GHzbandwidth with 31.25 kHz frequency resolution
•Continuous spectroscopic processing with no dead time