Page 380 - A Practical Guide from Design Planning to Manufacturing

350 Chapter Eleven

Imagine tests of the same processor in a four-processor system find that
when divide by zero errors are reported on all four processors in the
same clock cycle, the system may crash. This is also clearly a design flaw,
and the symptom is much more severe. However, this circumstance
would probably never be encountered in a real system. Divide by zero
errors are very rare and for four to occur simultaneously is astronomi-
cally unlikely. If system crashes caused by software errors are orders of
magnitude more likely than encountering a hardware bug, fixing the
hardware bug will provide no real gain in system reliability. For this
reason, processor designs will always have errata.
Once silicon debug has found and corrected all the design bugs, which
are severe enough to prevent shipping, the last step before being sold
is developing a silicon test program, which will separate out manufac-
turing defects.

Silicon Test
Only a very small percentage of parts ever go through the detailed
analysis of silicon debug, but every part sold must go through silicon test.
Manufacturing defects prevent some die from functioning correctly, and
there will always be manufacturing defects. If the process could be
developed to the point where there were no more defects, the feature
sizes would be shrunk to reduce cost and increase performance until
manufacturing defects began to appear again. The most profitable point
is always pushing the process to where some percentage of die fail, but
this requires a reliable way of separating good die from bad. In addition
to causing some outright defects, manufacturing variation will cause
some die to run at lower frequencies or higher power than others. These
parameters must be measured as well to make sure the processor will
not only be logically correct but will also meet its power and perform-
ance specifications. Silicon test is the process of developing and apply-
ing a test program to guarantee the quality of the processors shipped.
An example silicon test flow is shown in Fig. 11-6.
Silicon test begins even before the processors have completed fabri-
cation. Because it can take weeks to process a wafer, it is important to
be able to check the quality of the wafer before it is finished. Electrical
test (E-test) measures test structures on a partially processed wafer.
The processors themselves cannot be tested because they are incomplete,
but test structures allow parameters such as transistor on-current, off-
current, and wire resistance to be measured. These test structures are
built into the die or made in the scribe lines in between the die, which
will later be destroyed when the wafer is cut into individual parts.
The results of E-test are used as feedback to the fab. If measured E-test
parameters begin to fall outside the expected range, fab work may be