BIRD61 - Input Model

To IBIS Committee:

I have been following the BIRD61 discussions. At this time, we are
still looking for a solution to simulate delay variations as a function
of actual input waveforms Vin(t).

The reference model that captures delay variations is the Spice model
itself. However, for business reasons (versus technical reasons), this
model may not be available.

BIRD61 attempts to work around this limitation by providing delay
data (that may be derived from Spice Model simulation). The simulator
then is supposed to use this data to derive an equivalent model of
delay changes that can be applied for an arbitary Vin(t).

BIRD61 proposes to tabulate delays as a function of IDEAL input ramp
parameters: Start_point, End_point, and Slope.

Two general approaches can be taken to apply this data:

(1) Develop a method to characterize an arbitrary Vin(t) into table
parameters and use the delay tables directly.

(2) Use the table data to develop model of the input from which Vin(t)
can be transformed (through simulation or tranformation) and the resulting
delay can be derived. The authors intended that this approach be taken.

Approach (1) is attractive. It is possible to provide a table as a
function of all three parameters that captures the changes over the
ranges of interest. The table ranges of interest would need to include
the min and max values of all of the parameters. Other intermediate
values might be needed. However. the amount of data would depend upon
how accurate the delays are when such data is extracted using interpolated
values of Start_point, End_point, and Slope.

The problem is to characterize an "arbitrary" input into the "effective"
Start_point, End_point, and Slope. While there is no perfect process,
some defendable approximation approach might stated such as:

  Start_point = Min point,
  End_point = Max point,
  Slope = Slope from Min to Max, but derated by some known manner.

The problem is to find a good derating method. Some approaches might be to
assume a fixed factor, use the (set of) test waveforms to derive a factor,
or use some mathematically derived effective slope.

The recommended derating method should be one that we agree upon. The
model provider then has enough information to populate the table based
on ideal ramp inputs (that do not need derating) and also to test the
accuracy of applying the derating method on real waveforms.

Approach (2) is attractive for its simplicity. However, in order for
Approach (2) to be practical, we really do need use the same input model
because:

  Simulators need to produce consistent results.

  Model providers need to know how the data is used in order to produce
    a sufficient amount of tabular data and to also check the validity
    of the calculated delay adjustment against the reference model delay
    simulation.

So far, the proposers of BIRD61 have investigated, without success, some
equation based input models.

These models could be of the following forms:

  (a) Vin(t) transforms into a delta delay value directly,

  (b) Vin(t) transforms into a Vout(t) from which the delta delay value is
       calculated.

While I can speculate on a number of approaches, I really do not know
where to begin. Approach (a) may be to capture as part of the equation the
delay as a function of voltage (under and overdrive effects extracted from
the real Vin(t)) and also adjustments as a function of slope (derived in
some manner from Vin(t)).

Approach (b) is the mathematical equivalent of the complete Spice model
itself (or a simplification). In the extreme case, a (large) set of
mathematical non-linear equations could be provided that are the Spice
equations with the given process parameters. These equations would
relate Vin(t) and Iin(t) of the input node to Vout(t) and Iout(t) of the
output node. No table data is needed. The given equations are applied
directly through Spice simulation.

If a simplified set of equations are presumed, table data is needed to
fit the coefficients or parameters. However, for consistency, we all
need to use the same set of equations.

Also, the question arises whether a generic Spice input structure might
be presumed (implying only one set of equations). One problem may be
that the format of the set of equations will vary with different input
structures.

I think we should continue to pursue Approach (2) and may actually discover
a solution. However, we need some break-through information such as a good
starting point for an equation structure or a good input model (e.g.,
operational amplifier with feedback) to investigate this further.

In the mean time, I think Approach (1) is more promising.

Bob Ross
Mentor Graphics
Received on Tue Sep 14 17:27:21 1999