Browse Prior Art Database

Parameter to Assess Fault Clustering in Chips

IP.com Disclosure Number: IPCOM000102501D
Original Publication Date: 1990-Nov-01
Included in the Prior Art Database: 2005-Mar-17
Document File: 3 page(s) / 105K

Publishing Venue

IBM

Related People

Guedj, D: AUTHOR [+3]

Abstract

In a semiconductor line, one typically has access to test data giving the number of good chips per wafer (or lot) as well as their position on the wafer. If redundancy is not used, one simply knows whether a chip has zero fault or contains one or more faults. This information by itself is insufficient to uniquely find out the total number of faults and their distribution, and, therefore, to deduce a yield model.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 51% of the total text.

Parameter to Assess Fault Clustering in Chips

       In a semiconductor line, one typically has access to test
data giving the number of good chips per wafer (or lot) as well as
their position on the wafer. If redundancy is not used, one simply
knows whether a chip has zero fault or contains one or more faults.
This information by itself is insufficient to uniquely find out the
total number of faults and their distribution, and, therefore, to
deduce a yield model.

      The same chip yield can be achieved for vastly different number
of faults, depending upon whether faults are randomly scattered over
the surface or clustered together. In order to illustrate this, we
can choose a square wafer containing 225 chips (a square array of
15x15) among which 50 chips are fault-free. The overall wafer yield
is therefore 22%. If faults are assumed to be randomly distributed on
the chip, an average density of 1.5 faults per chip (or a total of
338 faults per wafer) will give a chip yield of .22. Ten faults per
chip ( 2250 faults per wafer) are needed to achieve an equivalent
yield with a clustering factor  of 0.49. We will use these two
examples in the following discussion to illustrate the difference
between random and clustered faults.

      A typical way to analyze clustering is the so-called 'block
method', in which the yield of blocks of 2, 3, 4... chips is computed
as a function of block size. A non-linear regression is then used to
obtain the three parameters YO, and needed to fit with a negative
binomial law. If we use this approach for our example, we obtain the
following table:
 Chips per block              Random                Cluster
         1                     .2222                 .2222
         2                     .0548                 .1452
         3                     .0102                 .0923
         4                     .0054                 .0683
         5                                           .0394
         6                                           .0361

      Yield of blocks of 1 2 3... chips are given as a function of
block size for randomly distributed (g = 1.5 faults/chip) or
clustered faults (g = 10 faults/chip , a = 0.49).

      The yield of blocks of chips grouped by 1,2,3... as a function
of the block size for random (solid) or clustered faults...