Spectral Built-In Self-Testing for Mixed-Signal Systems-in-a-Package
Abstract: In this research the PIs propose a new application of the theory of signals and systems to testing digital systems-in-a-package (SIPs). Digital input signals are considered as a set of time-varying waveforms, characterized by a correlation matrix. The elements of this matrix are the auto-correlation and cross-correlation coefficients, determined from inputs that have known fault and error detection properties. The auto-correlation indicates how a signal resembles its own past, and the cross-correlation indicates how a signal resembles other signals.
The PIs propose a completely integrated built-in self-testing (BIST) system for huge digital hardware systems. Rather than inserting a full-scan chain and using combinational BIST, digital circuit spectral analysis techniques is used to determine which flip-flops (memory elements) to scan to simplify the testing. Subsequent topologically analysis of the digital part of the system is made to determine its necessary spectral testing frequencies in partial-scan mode at each primary input. Finally, one inserts low-overhead spectral BIST hardware to generate these frequencies and uses a novel response compacter for BIST, which calculates the cross-correlation between various primary outputs as a signature. Spectral BIST uses less hardware than the conventional BIST system, covers 8.4% more of the faults in sequential mode, but requires less than 10% of the patterns.
The benefits of these concepts would be drastically shorter test pattern sequences, compared with present-day BIST, but high fault coverage. This leads to economic benefits of a large power reduction in test mode, and shorter test time. Other benefits are simplified testing hardware and lower hardware overhead. We will create a spectral analysis method to insert the testing hardware into the sequential circuit to raise its fault coverage. The PIs plan to topologically analyze sequential circuits to determine the predominant spectra of their input responses. This is expected to accelerate spectral analysis, since one can avoid analyzing parts of the spectrum that the circuit could never generate.