The frequency spectrum of today’s high speed digital signals extends into the microwave bands. Combating the effects that microwave engineers love (and which digital engineers dodged in college!) determines the success or failure of projects from telecoms and data center infrastructure, servers, PCs, and consumer electronics. In fact, for typical rise and fall times, the knee frequency (where the power starts to drop off rapidly) is 2.5 times the bit rate. At 10 gigabit/s you have to worry about frequency components at 25 GHz or even higher.
Microwave savvy engineers know that an impedance discontinuity whose length is on the order of a wavelength or larger gives rise to unwanted reflections. The wavelength at 25 GHz in FR4 printed circuit board is smaller than your pinky, about a quarter inch (~6 mm). Besides reflections, the interconnection path can introduce other impairments like distortion and crosstalk. The I/O circuits themselves inject finite jitter in the signal. (Jitter is a term used by high speed digital engineers and is closely related to the microwave concept of phase noise.)
The I/O circuitry of today’s high-speed digital ICs make use of sophisticated signal processing techniques in order to mitigate these impairments that occur due to distortion, impedance mismatch, reflections, and crosstalk from the interconnection path. Algorithms familiar to wireless baseband engineers -- pre-emphasis, adaptive equalization, and clock-data recovery -- are being repurposed to ensure chip-to-chip signal integrity. The state-of-the-art is now so sophisticated that the circuitry besides each bond pad can be considered part of a “mini-communication system.”
To use these new capabilities to their fullest extent, signal integrity engineers at the OEMs require accurate models of the ICs that they purchase. They plug these models into a system simulation in their EDA tools, where they act like “executable datasheets” for each IC in the system. Simulation is preferred over cut-and-try prototyping because it’s quicker, less expensive and it gives you more insight: you can plot waveforms at any point, no microprobes required.
But the details of the signal processing techniques and implementations are valuable intellectual property to the IC vendors. As a result, the information is first protected with encryption and generally only released under a confidentiality agreement. However, much like the music industry with its fragmented, proprietary digital-rights-management (DRM) methods for copyright protection, the EDA industry has failed to create an open encryption standard1. As a result, balkanization occurred: IC vendors were forced to encrypt and re-encrypt the same SPICE netlist with multiple proprietary encryption keys -- one from each EDA vendor -- and verify and support each version of the model.
The Algorithmic Modeling application programming Interface (AMI) extension to the Input/Output Buffer Information Specification (IBIS) standard circumvents this issue because models conforming to this API are Windows DLLs (or, equivalently, shared objects for Linux) compiled from behavioral source code. AMI is new feature in the latest version (5.0) of the larger IBIS standard 2. IC vendors can share these models without proprietary encryption. One model can run in any EDA tool that supports the standard.
An additional benefit is that compiled behavior-level models run a lot faster than the equivalent transistor-level netlist. With faster models, the OEM can complete a bunch of “what if” design space exploration simulations and very rapidly converge on the optimum set of design parameters such as connector type, trace and via sizes, equalizer tap settings, and so on.
Using IBIS AMI in the signal integrity workflow
Sounds good in theory, but how do you actually use the models once your IC vendor gives you them? Because AMI is an extension of IBIS, look for the IBIS model library in your EDA tool, drop the component into the schematic like so:
Figure 1: IBIS AMI components in a circuit simulator. The example shown here is a beta version of the channel simulator in Agilent Advanced Design System (ADS) 2010. Notice that the simulation controller in this case isn’t a SPICE-like transient simulator: it’s a much faster time-domain simulator called a channel simulator. It employs a short transient simulation to probe the channel with a step function.
…then open the component and go to the AMI tab:
Figure 2: Double clicking on an IBIS AMI component in the ADS schematic opens its dialog box. One tab (show here) lets you specify the appropriate AMI file that your IC vendor has provided.
If instead of just a waveform you want an eye diagram, you attach the Eye Probe component before running your simulation. Hey presto, the eye diagram, BER contours, and bathtub curves are generated in seconds. You can play around with the settings manually or the EDA tool’s optimizer automatically tune the circuit for maximum eye opening.
The example plots in Figure show the effect of the receiver equalizer.
Figure 3: IBIS AMI receiver model, a) Eye diagram before equalization, b) Same IBIS AMI receiver model, eye diagram after equalization. c) Timing bathtub curves (blue and red traces are before and after equalization, respectively) d) Voltage bathtub curves (again blue and red traces are before and after equalization, respectively)
Because ADS is an integrated platform, you can combine IBIS AMI models not only with traditional SPICE-like lumped elements (R, L, C, semiconductors, etc.) but also distributed elements like transmission lines. The transmission line models of your channel can be generic pre-layout multi-layer models from a 2D EM simulator or can be specific post-layout ones you generate from a 3D full-wave EM simulator like ADS Momentum. Or you can import measured data such as a multi-port s-parameter file. In any case ADS Transient Convolution uses a patented algorithm to ensure models created from frequency-domain data satisfy the Kramers-Kronig3 relationship and are thus causal in the time-domain.
In summary, IBIS AMI models help signal integrity engineers build up a complete and accurate system model which can be optimized in simulation, rather than resorting to expensive and time consuming cut-and-try prototyping. Agilent is investing in streamlining both the creation and application of AMI models in order to accelerate adoption of this new standard, to enable faster signal integrity design and verification.
Footnotes
1. But there is an effort to create one called IEEE P1735
2. The IBIS project began in the 1990s and version 1.0 appeared in April 1993. Version 2.1 was formally ratified as ANSI/EIA-656 on December 13, 1995. Versions of IBIS before 5.0 modeled only the analog part of the I/O characteristic such as the transmitter drivers and receivers.
3. Understanding the Kramers-Kronig Relation Using A Pictorial Proof (PDF) http://www.agilent.com/find/eesof-kramers-kronig