What is guard band?
04/01/1997
What is guard band?
Tricia Justice, Credence Systems Corporation, Fremont, California
Guard bands enable test limits to be adjusted to allow for the worst-case measurement error. Guard banding guarantees that the device under test (DUT) is within specifications, even if the measurement equipment is operating at its lowest level of accuracy. This article describes the theory behind guard banding and explains how to guard band measurement results based on the accuracy specification of the measurement equipment.
In the cases discussed here, the amount of guard band required is derived solely from the accuracy specifications of the measurement equipment. Other factors, such as temperature fluctuations, may also be taken into account when determining the guard band.
To illustrate, consider the following timeline:
The three values shown are the DUT timing specifications for a given timing parameter. The DUT specification sheet guarantees that this timing parameter is never less than tmin or greater than tmax, and the typical value is ttyp. If the test equipment used to measure this parameter were perfect, the passing region for this test would be:
tmin = tmeas = tmax
where tmeas = the value measured by the tester.
Unfortunately, no test equipment is perfect; therefore, a guard band must be added to the test`s passing region to guarantee that the DUT meets its published specification. Assuming the test equipment has a maximum error of ?terror, the illustration below shows the guard band regions (shaded) for the measurement:
where tmin` = tmin + terror and tmax` = tmax - terror
and the new passing region is
tmin` = tmeas = tmax`
An example shows why the guard band is needed to ensure the device`s timing specification. Assume the tester is off by the maximum amount (+terror) and, without guard banding, the tester measures tmeas = tmin. This would be a passing result. Because the tester measurement is off by + terror, the actual result is tmin-terror, which should fail. Without guard banding, this device would pass when it should fail.
Conversely, when guard banding is used, some devices that should pass will fail. For example, assume that the tester error is at the minimum (error = zero) when tmeas is measured and that the guard band has been applied to the limits so that the lower limit is tmin`: the DUT could fail if the tester makes a measurement where tmeas = tmin. Since the actual error of this measurement is zero and not the worst case, terror, this device should pass. From the standpoint of yield, then, the smaller the guard band the better.
Calculating the guard band
Derivation of the guard band is primarily dependent on the number of tester resources. It is assumed that all other tester resources not involved in the measurement have ample margin and need not be considered when determining the guard band.
First order measurements. Measurements involving a single tester resource will be referred to as first order measurements. First order measurements are typically DC tests, such as those measuring a voltage or current value. It is relatively easy to figure out the guard band for first order tests.
The accuracy specifications for the tester`s measurement resources are published by the tester manufacturer. Theoretically, the accuracy specification is stated as ? xerror for the tester resource making the measurement.
For example: xerror for Idd measurements is the accuracy specification of the power supply current monitor; xerror for leakage tests is the accuracy specification for the PMU current monitor; and, xerror for voltage measurements is the accuracy specification for the PMU voltage monitor. Once xerror is calculated for these measurements, the following equations apply:
xmin` = xmin + xerror
xmax` = xmax - xerror
where xmin` and xmax` are the test limits and xmin and xmax are the published device specifications.
Guard banding can also be applied to AC or DC functional tests, known as go-no-go tests. In a go-no-go test, a given resource is programmed to the minimum or maximum device specification value and then tested functionally to see if the DUT operates at the limit. In this case, we are programming the hardware to the limits, as opposed to comparing a measurement result to the limits.
An example of a first order go-no-go test is a Vil/Vih (low and high input voltage thresholds) functional test. Usually, the Vil specification is designated as a maximum limit and the Vih is specified as a minimum limit. Using the above equations, the high and low driver levels can be determined. In this case, ?xerror is the accuracy of the driver rails: xmin is the published Vih specification and xmax is the published Vil specification. The terms xmin` and xmax` are the programmed Vih and Vil levels, respectively.
Second order measurements. Measurements involving two tester resources will be referred to as second order measurements. Examples include propagation delay (one force edge and one compare strobe); setup and hold time measurements (two force edges); and pulse width measurements (two compare strobes).
Using two tester resources, the accuracies of both measurements must be considered when determining the guard band. Again, it is assumed that all other device parameters have a very large margin and guard bands will not be considered for them.
In this model, tdiff is the time difference between two edges. Note, this could be setup time, hold time or propagation delay. The task is to find the guard banded limits, tmin` and tmax`, given that
tdiff = tedge2 - tedge1
where tedge1 and tedge2 are the programmed or measured edge
positions
and tmin = tdiff = tmax
where tmin is the specification for the minimum tdiff, tmax is the specification for the maximum tdiff, and ?terror is the accuracy spec for each edge
First we will find the guard band, or the worst-case accuracy, for the lower limit of the measurement (see below):
As shown, the worst-case accuracy for the minimum limit would be tedge1 - terror along with tedge2 + terror. Combining these errors causes a result that is larger than it should be, so the lower limit may pass when it should be failing. Substituting these values into the equations gives the following:
tdiff` = (tedge2 + terror) - (tedge1 - terror)
= tedge2 - tedge1 + 2terror
= tdiff + 2terror
Let tdiff = tmin
so that tmin` = tmin + 2terror
Therefore, the lower limit of this measurement must be increased by +2terror to ensure that bad devices do not pass due to the inaccuracy of the measurement equipment.
Now we will find the guard band for the upper limit of the measurement. The worst-case scenario for the upper limit is:
The worst-case accuracy for the maximum limit occurs when tedge1 + terror along with tedge2 -terror. Combined, these errors yield a result that is smaller than it should be. Therefore, the upper limit may pass when it should be failing. Substituting these values into the equations gives the following:
tdiff` = (tedge2 - terror) - (tedge1 + terror)
= tedge2 - tedge1 - 2terror
= tdiff - 2terror
Let tdiff = tmax
so that tmax` = tmax - 2terror
Therefore, the lower limit of this measurement must be decreased by -2terror to guard against bad devices passing due to the measurement equipment`s inaccuracy. In summary, the new passing region is:
tmin` = tdiff = tmax`
where tmin` = tmin + 2terror
tmax` = tmax - 2terror
Compare this to the guard band for first order measurements. In these simplified examples, the guard band for a second order measurement takes into account the errors of two edges, while the guard band for a first order measurement includes only one.
In a second order measurement, the accuracy specification for the two edges are not always equal. If they are not equal, the term 2terror can be replaced by the sum of the errors of each edge.
Conclusion
Guard banding can account for inaccuracies in the measurement procedure and prevent defective parts from being approved.
For more information, contact Tricia Justice at: Credence Systems Corp., 215 Fourier Ave., Fremont, CA 94539; ph 510/657-7400, fax 510/623-2560, e-mail [email protected].