Backchannel Modeling and Simulation Using Recent Enhancements to the IBIS Standard

Backchannel Modeling and Simulation Using Recent Enhancements to the IBIS Standard By Ken Willis, Product Engineering Architect; Ambrish Varma, Senior Principal Software Engineer; Dr. Kumar Keshavan, Senior Software Architect; Cadence Recent enhancements to the upcoming IBIS standard now support backchannel training, enabling IBIS- AMI models to emulate this real-world SerDes behavior. AMI modelers now can incorporate backchannel algorithms into their IBIS-AMI models, automating the optimization of transmitter and receiver equalization settings in the same manner as their actual SerDes hardware devices. This saves system designers significant time by avoiding a multitude of computationally intensive sweeping to determine optimum equalization settings for their link, while at the same time yielding more realistic and higher quality results that are more consistent with the hardware they seek to model. Contents Backchannel Basics...1 Backchannel Enhancements to the IBIS Standard...2 How Backchannel Simulations Work...3 Simulation Testbench Setup...6 Sweep Results...7 DFE Adaptation...11 Conclusion...13 References...13 Backchannel Basics Backchannel capability, also called link training in the context of serial links, is the ability of the serial link receiver (RX) to automatically tune the equalization settings of the serial link transmitter (TX) to optimize signal integrity (SI) and bit error rate (BER). This capability exists in popular serial link standards such as InfiniBand, Fibre Channel, SAS, 10GBASE-KR, and PCI Express 3.0 and 4.0 [1], [2]. Backchannel training enables the equalization settings of the RX and TX to be optimized in combination. This produces better BER margins than just the TX or RX being optimized individually, which is critical as margins tighten up at higher data rates. This also enables simulation to better mimic the behavior of the actual hardware, and enables engineers to make better design decisions based on their simulation results. In hardware, the TX and RX enter a training mode in which the TX is open to suggestion from the RX as to how to adjust its feed-forward equalization (FFE) settings. IN Z -1 Z -1 W 0 W 1 + OUT Figure 1: Feed-forward equalization

FFE is configured with several taps, or boost drivers, that help emphasize the high frequency (or de-emphasize the low frequency) aspects of the transmitted signal. These taps have coefficients to control their drive strength, and these coefficients have practical limits that must be observed. During backchannel training, the TX sends a test pattern to the RX. The RX then evaluates the signal quality of the transmitted signals and communicates back through the physical channel (i.e., backchannel) to the TX, telling it how to adjust its FFE settings. The settings are adjusted, a new pattern sent to the RX for evaluation, and the process continues until either the RX is satisfied with the signal quality, or the time or cycle limit is reached for the training process. When the TX FFE settings are locked in, then the actual data transmission occurs. This process is illustrated in the figure below. Start TX Sends Test Pattern TX Adjusts EQ End RX Evaluates Test Pattern Standard Serial Link SI Flow Training Phase Complete Yes SI OK? No RX Recommends EQ Adjustments Figure 2: Backchannel process Backchannel Enhancements to the IBIS Standard I/O buffer information specification, or IBIS [3], was ratified in 1993 to serve as a device modeling standard to enable SI simulation. In 2007, an algorithmic modeling interface (AMI) section was added to the specification to accommodate the equalization functionality found in modern serializer/deserializer (SerDes) devices, which brought a whole new level of simulation functionality to the industry. Recently, a buffer issue resolution document (BIRD 147) (4) that adds backchannel support was approved to go into the next revision of the IBIS specification. Cadence has implemented early support for this functionality based on both internal and external customer requests. The IBIS implementation of backchannel support is defined as a simple file interface, and consists of the addition of five new Reserved _ Parameters to go into the.ami files of the associated IBIS-AMI models. The following Reserved _ Parameters should be present in both TX and RX models that support backchannel training: BCI _ Protocol. The name of the standard protocol being used. Both the TX and RX need to be pointing to the same protocol for link training to be enabled. A model can support more than a single protocol. In the current version of the specification, only private protocols can be defined. There is provision in the specification to make IBIS-approved protocols public for backchannel training that the models may support in the future. BCI _ ID. The prefix to be used for the file passed between the TX and RX. The rules for BCI _ ID and associated files are protocol-dependent and described in the BCI _ Protocol. The private protocol used for this study requires that the TX and RX read and write the same file with the BCI _ ID of cdns _ bci. BCI _ State. The state of the model. The acceptable parameter values are off, training, converged, failed, and error. Both the TX and RX can update the BCI _ State parameter. The EDA tool can terminate training and continue with the traditional channel simulation if the BCI _ State is converged. The receiver AMI model requires two additional parameters to support backchannel: BCI _ Message _ Interval _ UI. Suggested AMI _ GetWave block size to use for waveform passing between the TX and RX AMI models. BCI _ Training _ UI. Maximum number of UIs to evaluate for backchannel training. The EDA tool uses this parameter value to terminate link training if both the models have not come to a mutual state of convergence. This is important to limit any runaway routines or infinite loops during training. www.cadence.com 2

The IBIS-AMI models for both the TX and the RX must contain the associated AMI Reserved _ Parameters to be backchannel-capable. How Backchannel Simulations Work For backchannel training to be enabled in an IBIS-compliant channel simulator, a few simple prerequisites must be met: Both the TX and RX have IBIS-AMI models assigned In those IBIS-AMI models, both TX and RX models have the BCI _ State set for Training Both TX and RX models have identical BCI _ Protocol settings When the above conditions are in place, backchannel training is enabled, allowing the TX FFE settings to be tuned and set. When training is complete, channel simulation runs with the FFE settings in place, and the data transmission is simulated. The sections below describe the sequence of events in a five-step process that occurs at simulation time when a backchannel is initiated, using the PCI Express 4.0 model that Cadence developed with the new IBIS-based backchannel capability as an example. To enable an IBIS-AMI model with backchannel training capability, code modifications are required to perform each of these five steps. 1. The transmitter initiates the process With backchannel enabled, the TX sends a pattern of 1s and 0s to the receiver through the defined channel as a test pattern. This test pattern can be pre-defined by the BCI _ Protocol or can be run with any PRBS stimulus pattern set up in the simulation tool. In the models used for this study, the TX accepts the PRBS pattern generated by the simulation tool (Cadence Sigrity SystemSI technology) and does not require any specific pattern for link training. Next, the TX communicates to the RX what its capabilities are. This is done through a file interface. The TX writes out this file initially, using the BCI _ ID prefix defined in both AMI models. This file is read by the RX model to find out the status of the TX taps. The RX model opens the file because both models support the same BCI _ Protocol, and the protocol determines the rules for reading and writing the file and its contents. Following is an example of the initial BCI file written out by the Cadence TX AMI model: (amitx (tapincdec (-1 0) (0 0 ) (1 0 ))) This simple two-column format is set up with number pairs in each line for <tap _ id><tap _ state>, which indicates the following: Tap Number What it Means tap _ id 1 Precursor tap tap _ id 0 Main tap tap _ id 1 Post-cursor post tap tap _ state 0 the tap is open for adjustment tap _ state 1 the tap is at its maximum limit tap _ state 1 the tap is at its minimum limit Table 1: tap_id and tap_state The TX writes out a text file that tells the RX how many taps it has, the tap configuration (i.e., how many pre-cursor and post-cursor taps exist), and whether they are open for adjustment up or down in strength. In the example file shown above, all three taps are open for adjustment in either direction. www.cadence.com 3

The rules for tap adjustment must be documented with the TX AMI model as the RX model needs to know the limits of the TX model. (In this case, as both the TX and RX models were Cadence AMI models, this was not a concern.) 2. The receiver evaluates the waveforms The receiver then evaluates the signal quality of the waveforms it receives. There is no pre-defined method to do this; the RX AMI model is simply a black box. It accepts the received waveforms, applies its AMI equalization functionality, outputs those equalized waveforms, and then evaluates the resulting signal quality using its own algorithm, however it is defined. One example algorithm that SerDes receivers may apply is based on error sampling (figure 3). The receiver may monitor the median of the voltage distribution of the eye density, and can measure the delta between where the actual waveform voltages fall at the sampling point vs. the median value. That delta is considered the error value, and can be either positive, meaning above the median value, or negative, meaning it falls below the median value. Figure 3: Signal quality evaluation by the receiver The error sampling algorithm in the DFE corrects the post-cursor taps only in the RX, but the sampler has enough information from previous samples about the contribution to the overall error from the pre-cursor taps as well. It then uses this information to decide which direction the correction needs to be done, and conveys this information to the TX. This correction must be averaged over a period of time to provide stability and reduce dithering in the tap values. A typical sampler algorithm is based on the following equation: UP_DN n = ƒ n (ε 0,S n ) Where UP_DNn is the up- or down-correction for the n th tap; ε 0 is the error at the current location based on the median of the desired distribution (b0 in the figure above); S n is the state of the n th tap (e.g., 1 or 1 for NRZ). For a simple implementation of the function, the possible values of UP_DN are 1, 0 and 1, indicating to the TX to decrease, keep constant, or increase the n th tap value. This is just one method of determining the signal quality of the waveform processed by the RX. For backchannel training, each RX must analyze the waveform and determine whether the TX taps should be modified further to improve the quality of the signal. 3. The receiver requests TX FFE adjustments Based on the evaluation and knowing the current state of the TX s FFE settings, the RX then writes out the adjustments it wants the TX to make to its FFE settings for the next training transaction. In the models used for this study, the RX model does this by overwriting the same file it originally got from the TX, with a slightly different meaning for the second column. www.cadence.com 4

Here is an example of the initial BCI file written out by the Cadence RX AMI model: (amirx (tapincdec (-1 1) (0 0 ) (1-1 ))) Again, a simple two-column format is set up with number pairs in each line for <tap_id><tap_equalization_directive>, which indicates the following: Tap Number What it Means tap_id 1 Precursor tap tap_id 0 Main tap tap_id 1 Post-cursor post tap tap_equalization_directive 0 Make no adjustment tap_equalization_directive 1 Increase equalization tap_equalization_directive 1 Decrease equalization Table 2: tap_id and tap_equalization_directive The RX overwrites the original BCI file, indicating what FFE adjustments it wants the TX to make. Because both the TX and RX support the same BCI _ Protocol, the RX understands how much the TX tap values will go up or down when it requests a change in equalization. 4. The transmitter adjusts and re-transmits The TX reads the newly-written BCI file, adjusts its FFE tap settings, and sends a new pattern for evaluation by the RX. 5. The backchannel training completes This overall process described above occurs multiple times until one of two things happen: The RX writes out a BCI file that indicates no further adjustments are needed (i.e., all tap_equalization_ directives in the BCI file are written out as 0) and changes the Reserved _ Parameter BCI _ State to converged. The BCI _ Training _ UI limit is reached In either case, when one of the two conditions above are met, the backchannel training terminates, the FFE settings in the TX are locked down, and the full channel simulation is run by the simulation tool. Backchannel Simulation Example To study the impact of backchannel training on signal quality, the following testbench was constructed (figure 4). AMI AMI no BC AMI PCIe 4.0 Single Channel Topology AMI BC TX_Primary TP1 Channel TP2 RX_Primary Figure 4: Simulation testbench www.cadence.com 5

Simulation Testbench Setup The channel block was populated with a family of scattering parameters, or S-parameters, ranging in physical length from 2 to 40 inches, and representing insertion losses of 1.4dB to 21.7dB at 8GHz, which is the Nyquist frequency for PCI Express 4.0 data rates of 16Gbps. These S-parameters are plotted in the figure below. Figure 5: Channel S-parameters Two sets of IBIS-AMI models are included in the testbench. They are identical; they contain the same spec-level PCI Express 4.0 functionality. The one difference is that one set is enabled with IBIS-based backchannel functionality, while the other is not. Shown below are the step responses for the channel models in the topology. Figure 6: Channel step responses www.cadence.com 6

The associated impulse responses are also shown in the figure below. Figure 7: Channel impulse responses As can be observed from the previous plots, the shorter channels exhibit significant reflection behavior, while the longer channels are almost entirely dominated by loss, and the magnitude of the reflections are largely attenuated by the time they reach the receiver. Sweep Results Two sets of channel model sweeps were set up for the testbench. Both sets of sweeps had adaptive, 2-tap decision feedback equalization (DFE) at the receiver. For the first set of sweeps, the TX FFE was allowed to self-optimize its tap coefficients based on the impulse response of the channel. This employs a least mean square (LMS) algorithm to calculate the optimum tap coefficients to open up the eye for the given channel. This is done strictly from the standpoint of the TX, without any knowledge of the RX equalization functionality. This produced the following results: Figure 8: Sweep results, no backchannel For the second set of sweeps, backchannel is enabled. Rather than having the TX self-optimize its coefficients without any knowledge of the receiver, the RX is allowed to evaluate the training patterns sent through the channel by the TX, and direct the TX to make adjustments to its equalization. www.cadence.com 7

Figure 9: Sweep results, with backchannel enabled The results with backchannel enabled were universally better than without across the metrics of eye height, jitter, and normalized jitter and noise (NJN). NJN is a signal quality metric defined as the normalized jitter and noise, and is computed by dividing the noise and jitter regions within one unit interval (UI) by the total eye area (figure 10). Figure 10: NJN The plots below (figures 11-12) show the percentage improvement of the three metrics (eye height, width, and NJN) across the spectrum of channel insertion loss. % Improvement 45 40 35 30 25 20 15 10 5 0 5 10 15 20 25 IL (db) Eye height improvement with backchannel enabled Eye width improvement with backchannel enabled Figure 11: Eye height and width improvement vs. insertion loss www.cadence.com 8

% Improvement 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 0 5 10 15 20 25 IL (db) Figure 12: NJN improvement vs. insertion loss The improvement is substantial; all metrics improved for each case. To give a composite picture of the improvement, the percentage of improvement for each metric was averaged together for each case. That plot is shown below. 18.0 16.0 14.0 12.0 10.0 8.0 6.0 4.0 2.0 % Improvement 0 5 10 15 20 25 IL (db) Figure 13: Composite improvement vs. insertion loss The general trend shows that improvement from backchannel increases as the channels get more lossy and less reflective, up to the point at which the equalization starts to reach its limits in compensating for the loss. This makes some intuitive sense, as equalization is designed to compensate for channel loss, and is generally less effective in counteracting reflection. FFE Adaptation To understand more about the effect of backchannel training, we will examine one of these cases more closely. Looking at the 31-inch channel, we saw eye height improvement of 39%, jitter improvement of 9%, and NJN improvement of 4%. www.cadence.com 9

Figure 14: Eye contours for 31 channel, with (blue) and without (red) backchannel In both the self-optimized and backchannel cases, the FFE tap coefficients for this case began at the same values. But while the non-backchannel case kept those coefficients throughout the channel simulation, they changed significantly when backchannel training was enabled. These FFE tap coefficients are plotted below. Note that the pre- and post-cursor FFE tap coefficients are actually negative values, but are plotted here as absolute values to facilitate comparisons in the plot. FFE Tap Adaption Tap Coefficient 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 2.00 4.00 6.00 8.00 10.00 Time (usec) Pre (abs) Main (abs) Post (abs) Figure 15: Adapted FFE tap coefficients The magnitude of coefficients for both the pre- and post-cursor FFE taps decreased by about 34% and 13% respectively, meaning that the FFE was essentially turned down during the backchannel training process, leaving more of the equalization work to the RX and its DFE functionality. www.cadence.com 10

DFE Adaptation The DFE coefficients were plotted out for the same 31-inch channel case, with backchannel both off and on. Below is the plot of the DFE adaptation with backchannel enabled. Figure 16: DFE adaptation with backchannel on; tap 1 is blue and tap 2 is green Below is the result with backchannel turned off. Figure 17: DFE adaptation, with backchannel off; tap 1 is blue and tap 2 is green Note how with backchannel training enabled in Figure 16, the first DFE tap coefficient starts out low, increases to a higher level, then drops down to a lower level. This is an interesting behavior; initially the FFE post-cursor tap was actually over-equalizing the signal, and the first DFE tap was trying to compensate for it. Once it communicated to the FFE to turn down its post-cursor equalization, this DFE tap stabilized at a lower level. This contrasts with Figure 17, where the DFE tap goes high and stays high, fighting the FFE post-cursor over-equalization throughout the simulation. The final DFE coefficients reported for both cases are shown below. Case DFE tap 1 DFE tap 2 31 channel, BC off 0.0125 0.015 31 channel, BC on 0.0025 0.0175 % difference -80.0 16.7 Table 3: Final DFE coefficients with backchannel on and off For this case, while the FFE was turned down slightly, the first DFE tap was turned down significantly, and the second DFE tap was increased by almost 17%. www.cadence.com 11

Pre- and Post-Cursor Roles Regarding ISI The fundamental goal of equalization such as FFE and DFE is to help cancel inter-symbol interference, or ISI. The transmitter s FFE has the ability to cancel both pre- and post-cursor ISI, while the receiver s DFE, quite powerful with its adaptive capability, is limited to addressing post-cursor ISI. With this in mind, it was interesting to look at what would happen if we focused the FFE exclusively on pre-cursor ISI, and left all of the post-cursor ISI to the DFE. To enable this, the FFE was configured to enable only its pre-cursor tap, and the simulation for the 31-inch channel was re-run with backchannel enabled, allowing the RX to direct the pre-cursor setting of the TX. This improved the result even further than the original backchannel-enabled simulation, as shown below. Figure 18: Eye opening improvement with FFE focused on pre-cursor tap Eye height improved from 82mV to 95mV, eye jitter improved from 0.37 to 0.31UI, and NJN improved from 0.82 to 0.81. The pre-cursor FFE tap adaptation from this simulation is plotted below. Tap Coeffcient 0.2 0.18 0.16 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0 0.00 2.00 4.00 6.00 8.00 10.00 Time (usec) Figure 19: Pre-cursor only FFE tap adaptation In this case, the pre-cursor FFE tap adapted to a value of 0.13, as opposed to the value of 0.094 that occurred when both pre- and post-cursor taps were enabled. So the FFE is doing more pre-cursor equalization work than before. Figure 20 shows the plot of the DFE adaptation for this case. 12.00 www.cadence.com 12

Figure 20: Pre-cursor only DFE adaptation The first DFE tap adapted to a higher value of between 25mv and 30mV, about an order of magnitude higher than with the FFE post-cursor tap being actively used. This shows how the DFE is doing significantly more of the post-cursor equalization work. The second DFE tap then stabilizes at a lower value than before, around 12.5mV, vs. approximately 16mV in the previous case, a decrease of about 28%. This shows that equalization is doing a better job, and the second DFE tap essentially has less work to do. Having the FFE focus exclusively on the pre-cursor ISI and having the DFE do all the post-cursor equalization produced the best overall results in this case. Conclusion This paper provided an overview of backchannel training functionality in modern high-speed serial link interfaces and the recent enhancements to the IBIS specification that enable backchannel capability to be incorporated into channel simulations using IBIS-AMI models. Multiple simulation results were presented that incorporated backchannel training, all of which showed significant improvement over cases where the transmitter self-optimized its equalization settings, absent of communication with the receiver. The roles of pre- and post-cursor equalization by feed-forward and decision feedback equalizers were explored. Guidance on how to utilize the new IBIS functionality to incorporate backchannel training into IBIS-AMI models was also provided, to enable more accurate representation of the behavior of SerDes devices that have backchannel capability. References 1. Mohammad S. Mobin et al., TX back channel adaptation algorithm and protocol emulation with application to PCIe, SAS, FC, and 10GBASE-KR, DesignCon 2012. 2. PCI Express Base Specification Revision 4.0 Version 1.0 September 27, 2017. 3. http://ibis.org/ver6.1/ 4. http://ibis.org/birds/bird147.6.docx Cadence software, hardware and semiconductor IP enable electronic systems and semiconductor companies to create the innovative end products that are transforming the way people live, work, and play. The company s System Design Enablement strategy helps customers develop differentiated products from chips to boards to systems. www.cadence.com 2018 Cadence Design Systems, Inc. All rights reserved worldwide. Cadence, the Cadence logo, and the other Cadence marks found at www.cadence.com/go/trademarks are trademarks or registered trademarks of Cadence Design Systems, Inc. All other trademarks are the property of their respective owners. 10364 05/18 MC/DM/PDF