Mixed-signal alternate test and binning using digitally encoded signatures
- Autores: Álvaro Gómez Pau
- Directores de la Tesis: Joan Figueras Pamies (dir. tes.), Luz Balado (codir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2017
- Idioma: español
- Tribunal Calificador de la Tesis: Michel Renovell (presid.), Rosa Rodríguez Montañes (secret.), Adoración Rueda Rueda (voc.)
- Programa de doctorado: Programa Oficial de Doctorado en Ingeniería Electrónica
- Materias:
- Enlaces
- Tesis en acceso abierto en: TDX
- Resumen
Integrated circuit industry has always faced the necessity of testing and verifying the fabricated ICs in order to guarantee that no faulty circuits reach the market as well as the fabricated devices function within design specifications throughout their entire service life. In the current frame, the accomplishment of this objective entangles considerable economic implications and great technological challenges. Analog and mixed-signal circuit testing becomes an even more challenging endeavor since no efficient nor systematic procedures are available. Further, these challenges are strongly related to the process variability induced by the constant feature size miniaturization as well as by circuit operating conditions (i.e. PVTA variations). As opposed to the classic or specification based testing, alternate testing techniques have been lately adopted as a promising solution in such challenging scenario.
In this thesis, a novel methodology to encode the pass/fail regions in the alternate measurements space is proposed. The method relies on two phases. The training phase digitally encodes the test acceptance/rejection regions in the alternate measurements space using octrees. The testing phase corresponds to the actual production testing using the previously computed octree data structure. Such digital encoding approach presents a number of benefits, specially in terms of computational efficiency since the resulting octree structures are inherently sparse, what yields to fast training and testing times. Octrees have the advantage of being generalized to an arbitrary number of dimensions without any extra issue. Regarding such generalization potential, octrees can extend their clustering capabilities to more than two clusters, therefore facilitating the proposed IC quality binning approach without any incurred overhead. Further, the octree encoding algorithm is deterministic as it does not rely on a minimization algorithm as many of the state of the art clustering methods do. This is a desirable feature since the resulting encoding does not depend on convergence issues or the considered initial conditions. The simplistic recursive implementation, both for training and testing, make them affordable and easy to implement in stand alone systems as well as convenient for BIST applications.
Within the current "more than Moore" scenario, testing heterogeneous devices entangles a series of non trivial challenges which are even harder to cope than the ones existing in traditional CMOS circuits. This is so because of their inherently non electrical nature, depending on the exploited transduction principle, creates the need of complex stimuli generation. For the particular case of MEMS accelerometers, a mechanical excitation stimulus is needed in order to emulate in field operation. In this thesis, a method consisting on a variable speed vertical spinning wheel mounting the device under test has been proposed. The composition of output signals under such excitation conditions yields to an analog trace characterizing the defect level. The analog signature is used to test the device based on the octree encoding of the alternate measurements space or used for diagnosis purposes by means of a signature compaction and feature extraction procedure.
The selection and acquisition of analog signatures for alternate mixed-signal testing is an issue of major concern in the field. This thesis proposes a selection algorithm based on redundancy avoidance within the selected set of indirect measurements yet keeping the maximum information to perform the test decision. The application of such selection algorithm immediately translates into better test results.
