IBIS Accuracy Subcommittee Minutes 4/30/98

IBIS Accuracy Subcommittee Minutes

Thursday, April 30, 1998
Held at Digital Equipment's PKO3 building, 129 Parker St., Maynard, MA

Attendees

Greg Edlund, Digital Equipment (chair)
Fawn Engelmann, EMC
Bob Haller, Digital Equipment
Bruce Heilbrunn, Stratus Computer
Peter LaFlamme, Fairchild Semiconductor

Next meeting is Thursday, May 21, 1998 from 3:00 to 5:00 pm at Digital
Equipment's
PKO3 building, 129 Parker St., Maynard, MA.

MILESTONES

The test board milestone has slipped. Greg and Fawn report difficulty
making progress on the board due to busy schedules.

June, 1998: Post the IBIS Model Test Board to the web
September 18, 1998: Distribute the first draft of the IBIS Accuracy
Specification
October 1998: PCB West Conference, IBIS Summit, and first IBIS Class
February 1999: Present the IBIS Accuracy Specification at DesignCon99

EDITING

1. Scope

The subcommittee voted unanimously to adopt a limited scope for the IBIS
Accuracy Specification version 1 as defined by the following keywords
and sub-parameters:

[Package] R_pkg, L_pkg, C_pkg
[Pin] R_pin, L_pin, C_pin
[Model] C_comp, Input, Output, I/O, 3-state, Open_drain
[Pulldown]
[Pullup]
[GND_clamp]
[POWER_clamp]
[Ramp] dV/dt_r, dV/dt_f

On April 8, 1998, three members of the IBIS Accuracy Sub-Committee -
Bruce Heilbrunn, Bob Haller, and Greg Edlund - met with Paul Back of
3Com at Stratus Computer in Marlboro to discuss the scope of the IBIS
Accuracy Specification and appropriate test loads. Paul is a former
Stratus employee and has brought his experience with behavioral modeling
and simulation to the IBIS Users Group. Thanks to Bruce Heilbrunn and
Stratus for hosting the meeting.

Most of the differences in opinion were really differences in language.
At Stratus and 3Com, they use the term "model generation" to mean
extracting behavioral models from SPICE simulations. Their process uses
a resistive load whose value is tailored to the output impedance of the
driver. The term "validation" describes the process of comparing SPICE
and behavioral simulations for a variety of system-like test loads.
There is a significant difference between the Stratus process and IBIS
at this point. For "validation," Stratus likes to use a range of test
loads between 10 and 100 Ohms. IBIS is content with a typical 50 Ohm
load.

The "Digital camp" and the "Stratus camp" agreed on one thing: the set
of test loads that are required to cover a given driver's electrical
behavior is definitely dependent on the design of the driver. One can
think of this problem as analogous to the ASIC fault coverage problem;
the test engineer needs to come up with a set of test vectors that will
cover the known faults for a given ASIC technology. In the same manner,
the writers of the IBIS Accuracy Specification need to come up with a
set of measurements that cover the known electrical behavior of a given
I/O buffer family. Greg gave the example of the GTL output buffer,
which employs a feedback circuit to adjust the output edge rate as a
function of load impedance. Clearly, an IBIS datasheet for a GTL driver
that was verified at 50 Ohms may not be valid at other load impedances.

This is the compromise that we arrived at: the first release of the
IBIS Accuracy Specification will cover the I/O buffer families covered
by IBIS 1.1, namely push-pull, open-drain, and open-collector. Bruce
Heilbrunn has been concerned that the first incarnation of the IBIS
Accuracy Specification may not be usable for other more advanced I/O
buffer families. To address this concern, we propose adding a
disclaimer statement to the scope section of the document:

"The methods defined by this specification may be used with unspecified
I/O buffers families, although this specification makes no attempt to
insure coverage of their electrical behavior by the measurements and
metrics defined within this specification."

The next version of the IBIS Accuracy Specification will by necessity
address the thorny issue of advanced driver families and differences
among simulation engines.

Accuracy Levels

The subcommittee defined four levels of accuracy which are intended to
cover a wide range of user needs. Each accuracy level corresponds to
the availability of different kinds of test samples. The methods of
comparing test data against behavioral simulation data are dependent on
what kind of test sample is available.

We make reference below to two lab vs. simulation comparison metrics,
which will be clearly defined when we write section 4 of the
specification. In summary, metric 1 attempts to overlay two curves and
quantify the average distance between corresponding data points on the
curves in the x or y direction. Metric 2 measures whether one curve
fits within the envelope defined by two other curves that represent
process and temperature extremes.

Level 1 is the most accurate and is applicable when a vendor has the
capability of tweaking the process variables to achieve what he or she
believes to represent the typical, extreme fast, and extreme slow
conditions. He then runs behavioral simulations which utilize the
typ/min/max columns in the IBIS datasheet. In theory, the behavioral
waveforms should overlay the lab waveforms closely because the process
conditions are known, and metric 1 (average absolute percent error)
quantifies the discrepancies

Level 2 is useful when a vendor has samples that are known to have been
processed in a typical lot but does not have the capability or luxury of
processing corner silicon. Using metric 1, the vendor compares the lab
data with behavioral simulations based on only the typ column of the
IBIS datasheet. In the absence of corner silicon, the vendor compares
SPICE waveforms at fast and slow process corners with behavioral
simulations at min and max datasheet columns also using metric 1. The
fast and slow SPICE waveforms serve as a stand-in in for real lab data.
The vendor must rely on the capabilities of his device modeling
engineers to produce transistor model parameters that are in keeping
with the limits imposed by statistical process control. We expect level
2 to be the most commonly used.

Level 3 is useful when a vendor has accurate SPICE models available but
does not know the processing conditions of the lot from which his sample
part came, i.e. the part is a "random sample." He uses metric 1 to
compare behavioral and SPICE waveforms at all three corners but uses
metric 2 (envelope) to compare lab data to behavioral data. Metric 2
works well for IV curves and terminated loads, but it is not true that
the min and max waveforms will completely envelop all other waveforms in
the case of the unterminated transmission line.

Level 4 is useful when only a random sample is available. It may be
used by third-party model vendors who do not have access to SPICE
models. This level may also be valid for a customer who is seeking to
validate a semiconductor vendor's IBIS datasheet with no other data
available. It is probably not of much use to semiconductor vendors
since they have SPICE models available.

Note that there is no accuracy level defined for comparing IBIS and
SPICE data only. While metric 1 may be valuable for checking SPICE vs.
IBIS during model generation, this approach does not fit the definition
of accuracy, which involves comparing experiment (lab) and theory
(simulation). Lab measurements are an absolute necessity for insuring
the accuracy of an IBIS datasheet.

2. Required Measurements

The subcommittee approved of a first-cut at the required test loads for
push-pull drivers. These will appear in schematic form when we write
the specification. The required test loads for a push-pull driver are
as follows:

vendor-defined standard load
50 Ohms to Gnd (or 50 Ohm transmission line terminated by 50 Ohms to
Gnd)
50 Ohms to Vdd (or 50 Ohm transmission line terminated by 50 Ohms to
Vdd)
open-ended transmission line, length a function of rise time
driver driving receiver through a transmission line

The required test loads for an open-drain driver are as follows:

vendor-defined standard load
50 Ohms to Vtt (or 50 Ohm transmission line terminated by 50 Ohms to
Vtt)
open-ended transmission line terminated to Vtt at the near end, length a
function of rise time
driver driving receiver through a transmission line terminated to Vtt at
the near end

All of the push-pull test loads appear on the IBIS Accuracy Test Board.

At the next meeting we need to discuss the other required measurements:
IV curves and capacitance.

There was some discussion regarding the need for 10 Ohm and 100 Ohm
loads in addition to the 50 Ohm loads. Bob Haller argued that for the
drivers he's worked with, if correlation is good at 50 Ohms, it will be
good at other load impedances as well. Bruce Heilbrunn argued that
there are drivers that correlate well at 50 Ohms but not at the extremes
of load impedance. Bruce has the task of finding an example of an IBIS
1.1 driver that exhibits good correlation at 50 Ohms and poor
correlation at the extremes.

3. Measurement Techniques

We did not discuss measurement techniques at this meeting. The June and
July meetings will be dedicated to this topic.

4. Metrics

We did not discuss comparison metrics at this meeting. The August
meeting will be dedicated to this topic.

RELATED ACTIVITIES

IBIS Accuracy Test Board - Fawn and Greg have finished the schematics
for rev 2 of the accuracy test board and are now into layout. It looks
like it will be sometime in June before we have debugged hardware in
hand and are ready to release the design to the world. There have been
some people who expressed an interest in the board for training
purposes.

IBIS Developers Tool Kit - We received positive response from the
following vendors regarding their plans for making available a free or
low-cost simulation engine for model development purposes: HyperLynx,
Mentor/Interconnectix, Quad Design, and Cadence. Thanks to all for
their cooperation. We realize that it takes time to create such
software, especially when it is not a part of a product offering.

IBIS Cookbook - We did not report any progress on the cookbook. We
expect this activity to pick up when we get further into the measurement
techniques section.

HOMEWORK

Greg Edlund - Finish layout of the IBIS Accuracy Test Board.

Fawn Engelmann - Finish layout of the IBIS Accuracy Test Board.

Bob Haller - capture all required measurements in Word.

Bruce Heilbrunn - Identify a component that would serve as an example of
an IBIS 1.1 driver that exhibits good correlation at 50 Ohms and poor
correlation at the extremes. Demonstrate the magnitude of the
discrepancies between SPICE and behavioral simulations.

Peter LaFlamme - Identify and procure SPICE models for a component that
would serve as an example of an IBIS 1.1 open-drain driver.

----------
Greg Edlund, Principal Engineer
Server Product Development
Digital Equipment Corp.
129 Parker St. PKO3-1/20C
Maynard, MA 01754
(978) 493-4157 voice
(978) 493-0941 FAX
greg.edlund@digital.com
Received on Thu May 7 09:40:10 1998