Browse Prior Art Database

Enhanced Decode Detection

IP.com Disclosure Number: IPCOM000106932D
Original Publication Date: 1992-Jan-01
Included in the Prior Art Database: 2005-Mar-21
Document File: 4 page(s) / 264K

Publishing Venue

IBM

Related People

Canniff, ML: AUTHOR

Abstract

This article presents a program which analyzes array product Bit Fail Maps and separates out certain known fails for defects occurring in peripheral circuitry. Generic and accurate unlayering of address decode fault mechanisms is provided (Fig. 1). This unlayering becomes a key element of the Comprehensive Pattern Recognition Program. It deciphers the non-random full line failures versus randomly occurring point defects (Fig. 2). Unless these multiple line signatures are separated, there will exist a gross over count in the observed single line fault types (Fig. 3).

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 47% of the total text.

Enhanced Decode Detection

       This article presents a program which analyzes array
product Bit Fail Maps and separates out certain known fails for
defects occurring in peripheral circuitry. Generic and accurate
unlayering of address decode fault mechanisms is provided (Fig. 1).
This unlayering becomes a key element of the Comprehensive Pattern
Recognition Program.  It deciphers the non-random full line failures
versus randomly occurring point defects (Fig. 2).  Unless these
multiple line signatures are separated, there will exist a gross over
count in the observed single line fault types (Fig. 3).

      Decode Detection exploits the notion of equivalence classes and
partitions to detect symmetrical line patterns. These subjects are
related to set theory.  The program generates sets by using the
modulus operation and the set of positive integers I+.  This function
induces equivalent classes of sets of numbers when operated on I+.
For example, mod (X,4) will bucket any X into the values (0,1,2,3).
Therefore, mod (1,4) = 1 is equivalent to mod (25,4) = 1. Note that
any true symmetrical failing line pattern  will then have the same
modulo value (every 16th line out will possess the same mod (X, 16)).

      There exists an inherit ordering to partition refinement when
used with the modulo function.  The lowest refinement is, of course,
mod (X,1) = 0 for all integers. The highest would be mod (X,  I+  ).
The refinement can be thought of as a binary tree where (0)1 is the
root with (0)2 and (1)2 as children.  Each child has two descendants
and so on.  At each level, the modulo function is doubled (e.g., 1,2,
4,8,... or 1,5,10,20,...).  Now if we "pour" N numbers into the tree,
these numbers will continue to be sifted into higher modulo
refinements until there exists one number per mod value at some level
K in the tree.  When a true decode pattern exists, these line numbers
will "bunch" together along a single path in the tree for quite some
time. Actually, in our example above, the lines will follow the same
path until mod (X,16) (1og2 16 level).  There, they will split into
two even equivalence classes for mod(X,32) (1og2 32 level).  Any
random failing line will be sifted away from the pack and never be
considered (even though it may still be an element of (0)2 or (0)4,
e.g.).

      The main driver develops partitions and tree generation.  Since
there may exist failing lines which are adjacent to others, it
processes the largest group sizes to the smallest.  Grouping by size
is needed so that thin groups do not find decode members in the wide
groups.  This also enables large blocks to mask smaller decode
patterns.  The first act of partitioning will set up the root node of
the equivalence class tree.  This becomes (0)1 and contains all
groups.  The number of sons is determined by decode intervals.  The
first interval contains the powers of two: number of sons = 2; tree
levels are (1,2,4,8,16,....).  Nex...