News & Analysis
Synergy Needed For 10-Gbps Backplanes
Ken Lazaris-Brunner and Bharat Tailor, Gennum; John D'Ambrosia and Michael Fogg, Tyco Electronics
5/23/2002 4:21 AM EDT
The need to transmit serial gigabit data within optical networking equipment, such as routers and switches, is driving the need for increasing the aggregate bandwidth through a copper medium. Chip-to-chip communications on a printed-circuit board (PCB), card-to-card communications in a backplane environment, and copper cabling applications are all being re-visited in an effort to support the insatiable demand for bandwidth.
These applications all share a common overall structure, or system interconnect, which consists of a transmitter, a receiver, and the copper medium (also called the channel) between them. At gigabit speeds it is necessary for the designer to understand the entire system interconnect and the synergistic roles that the transmitter, receiver, and channel play.
In this paper, we'll layout a jointly developed card-to-card interconnect that delivers 10.7-Gbps data rates over 25 inches of FR4. We'll also look at an I/O cabling solution capable 6-m of cable at 10.7 Gbps.
Interconnect Evolution
Today, most interconnect architectures employed in optical networking designs deliver 2.5-Gbps data rates. When building these interconnects, designers typically select components in the channel based on their individual performance as opposed to their performance in a system architecture.
However, as backplane transmission speeds move from 2.5 Gbps to 10 Gbps and beyond, designers can no longer look at a component's performance capabilities as an island. Rather, designers must equally consider the individual performance capabilities of the components as well as the capabilities it exhibits when implemented in a system architecture.
The interaction between components results in a channel with unpredictable losses. It is this unpredictable nature that limits the effectiveness of using better performing materials or devices intended to compensate for the predictable losses between the driver and receiver. Thus, it is the underlying passive channel that ultimately limits the speed of operation and length for which it may run.
The channel is the electrical path between the driver and receiver. PCB traces and cabling are the two types of transmission media commonly used, while connectors and vias can be thought of as the transition mechanisms between segments of transmission media. Before considering the impact of the transition mechanisms, it is necessary to understand the transmission media and its performance capabilities.
It is easy to observe the significant difference in loss between different types of cables and PCB traces.1 Furthermore, while FR-4 is an industry staple, the inherent dielectric and skin effect losses of a trace in a PWB environment are perceived as a significant hurdle for multi-gigabit transmission for usable distances in a backplane system.
By using through-reflect-line (TRL) de-embedding, it has been shown that traces can deliver greater data capacity than originally thought. Simulations have shown that signal traces on FR-4 are capable of being driven with test equipment up to distances of 18 in. at 9.6 Gbps, while still maintaining a 23.8% eye-opening and 0.30 unit interval (UI) jitter.
The performance of other materials in the same configuration was also investigated. The test results are shown in Figure 1 and summarized in Table 1.
Table 1: PWB Materials Performance Summary
| Material | Eye Opening | UI Jitter |
|---|---|---|
| FR-4 | 23.8% | 0.30 |
| GeTek | 26.8% | 0.28 |
| Rogers 4350 | 42.6% | 0.20 |
| Arlon CLTE | 52.0% | 0.19 |
When two high-speed, backplane connectors were added to the same 18-in. trace, eyes for all systems were completely closed, regardless of the board material (Figure 2). Thus, it was shown that that the inherent dielectric and skin effect losses of a trace represent only one of the limiting factors in achieving multi-gigabit line rates. 2,3
Further understanding and optimization of the channel has been driven by examination of the PWB/connector interface. There are three aspects to this interface that need consideration. First, there are the physical constraints that are driven by the connector itself, such as signal and ground locations and hole or pad dimensions. Next, there are those issues that are influenced or limited by the connector, such as the differential trace geometry that may be routed through the connector pinfield, that influences trace width, spacing, and anti-pad dimensions. Finally, there is the interaction between the connector and the design of the system, which includes the overall board thickness and the relative position of the layer in relation to the overall board stackup at which a trace makes its electrical connection to the connector. All of these aspects combine to influence the overall performance of a given channel.4
The through-hole vias, which enable connections throughout the entire board stackup, are the heart of the problem. It is imperative to minimize the impedance discontinuity that a via will introduce in order to maximize the channel's throughput capability. While the discontinuity will be influenced by all of the various aspects of the PWB/connector interface, the layer connection in relation to the overall board stackup is a significant factor. Figure 3 illustrates the impact of the layer connection on the performance of the channel in the frequency domain.
Predictable and Unpredictable Losses
As Figure 3 points out, a backplane system top-layer connection has an unpredictable S21 profile, while the bottom layer connection is more predictable. Thus, the channel consists of two types of losses — predictable and unpredictable.
The predictable losses are those losses associated with the transmission media itself. These are mostly skin effect losses that increase with the square root of the frequency, and dielectric losses that increase linearly with frequency.
On the other hand, unpredictable losses, are caused by impedance discontinuities within the channel, which typically result from transitions between transmission media, such as in the connector/PWB interface. Furthermore these unpredictable losses can not be dealt with by merely dealing with the predictable-loss aspect of the system interconnect. Instead the impedance discontinuity itself must be dealt with in order to minimize the unpredictable nature of the channel.
This same phenomenon can be seen in cable assemblies. Predictable losses are related to the materials of the cable, while the unpredictable losses are related to impedance discontinuities in the cable assembly path.
The application of connectors (discontinuities) will vary the location in the frequency spectrum of high losses and reflections. Thus, the quality of the interconnect and the system design will affect the magnitude of the losses.
This implies a sound strategy for optimizing the channel in order to reduce unpredictable losses in the channel. By minimizing discontinuities, the resultant channel response is predictable and approaches that of the transmission media. Further improvements to the channel can be accomplished by dealing with the predictable losses through selection of better performing transmission media.
Obviously, an ideal transmission path would have no loss, no noise, and no delay. Using the techniques above, designers can develop a transmission media optimized for this performance. And, by developing optimized channel paths, designers will also make techniques like adaptive equalization more effective. To illustrate this point, we'll look at adaptive equalization and the impact a good transmission media has on equalization techniques.
Receive-End Adaptive Equalization
Once the channel has been optimized, it is necessary to select an active compensation method that is suitable for the channel. For years, design engineers have employed analog adaptive equalization techniques for digital TV signals transmitted over copper coaxial cable in TV studios using the SMPTE standards. This allows studios to take advantage of installed coaxial cable capacity. Current video equalizers are currently capable of receiving a signal over 150 meters of Belden 1694A coaxial cable at HDTV data rates of 1.485 Gbps. The same technique has been demonstrated to work over differential media such as differential cables, or differential traces.
In the discussion to follow, we'll discuss how the equalization method described above will work over cable. The methodology is exactly the same over traces, except the frequency-dependent attenuation curve has a slightly different shape.
The equalization algorithm works by apply the inverse transfer function of the frequency-dependent attenuation losses over the cable. This is expressed as:
where, ks, represents the skin losses, kd, represents the dielectric losses, and l is the length of cable.
Since the length term, l, is present in both frequency terms, the factor of el can be used as a multiplying factor. The shape of the transfer function is fixed, while the magnitude of the transfer function varies with the length. Hence, l is the only quantity that must be sensed by the adaptation algorithm. The shape of the transfer function is dominated by the skin effect losses, ks, at the frequencies of interest, and a single transfer function is adequate for all cable types.
As the cable length increases, the attenuation increases, slowing the rise/fall time of the transmitted signal. Eventually the rise/fall time of the signal decreases to the point that the signal does not reach its maximum (or minimum) value before the next symbol is transmitted, causing inter-symbol interference (ISI). At this point the eye starts to close, and the jitter starts to increase. Finally, the attenuation becomes so severe that the eye completely closes. Note that the equalizer is adaptive and will deliver an open eye in all cases up to its maximum gain as it compensates for the medium's attenuation characteristics.
It should be realized that an eye diagram is simply multiple traces overlaid one on top of the other. In reality, the signal is a single trace with distinguishable data pulses and the data information is still recoverable.
Transmitting the signal with a well-defined amplitude allows the equalizer to estimate the attenuation introduced by the cable by examining the spectral content in the received signal at a specific frequency. This estimate can then be used to apply the inverse transfer function to the received signal. This methodology allows usable signals to be recovered even when the eye is completely closed without any specific training sequences. Proper design of the attenuation estimation algorithm allows a large degree of data rate independence, and the equalizer can be used over a wide range of data rates.5
The same methodology can be applied to traces, merely substituting the trace frequency attenuation characteristics for the cable attenuation characteristics. As with cables, these are close enough among different trace designs and widths that the same frequency attenuation shape can be used for almost any trace.
Recent advances in adaptive equalizer technology can reduce the limitation on the transmitted signal amplitude so that the equalization can be made compatible with common mode logic (CML) or other highspeed standards. Given a reasonable channel, this equalization method offers several advantages:
- Wide range of trace lengths with no external adjustments
- Fast adaptation with no training sequences
- Reliable convergence even when no eye is present
- Compatible with both bi-level and multi-level signaling
Other active techniques such as pre-emphasis and adaptive linear transversal filters can be used. However, the discussions of these technologies are beyond the scope of this paper.
A Synergistic Solution
Above, we discussed designing a channel dominated by frequency-dependent attenuation and an equalization method designed to compensate for just this type of channel. While either approach alone will improve a given channel's performance, the synergy between the two methods will enable data rates and lengths that are useful in future high-speed backplanes and copper cable assemblies.
In the discussion to follow, we will show that a well-designed channel in conjunction with the adaptive equalization scheme presented here can significantly increase the length that can be achieved as well as decrease additive jitter.
Additive jitter after the equalizer is caused by three sources: 1) slight errors in the adaptation which cause errors in the correction curve, 2) differences in the correction curve and the actual trace attenuation, and 3) ripples in the channel attenuation due to reflections. With advanced design techniques, the first two sources of additive jitter can be reduced to less than 0.2 UI. The third source, however, can significantly impact the jitter performance of the equalizer.
Ripple in a channel is defined as the deviation from the trace-only attenuation. This is typically seen as a wavelike effect in the attenuation characteristic due to resonances from multiple reflections in the signal path. The equalization method assumes a trace-only attenuation characteristic due to the predictability of the reflections in different paths. Deviations from this curve cannot be compensated and the jitter will increase. For example, a channel with 6-dB peak-to-peak attenuation ripple (deviation from trace-only) at half the data rate will have an additional 0.25 UI of jitter above the initial 0.2 UI.
Note that the equalizer will adequately compensate for different trace lengths, essentially flattening the channel. These conclusions are therefore independent of trace length up to the maximum length the equalizer can handle. However, good channel design is often necessary to decrease channel to allow the equalizer to work to its maximum potential. There is also a maximum frequency limitation on applicable ripples.
Generally the frequencies of interest in the signal lie in the frequency range below half the data rate (i.e. less than 5 GHz for 10 Gbps data rate). Larger resonances at higher frequencies will have a smaller effect.
Previous experience has shown that the equalizer works well with predictable channels in the cable environment. As described in the section on channel characteristics, the channel can degrade this performance significantly. Figure 3 above showed how the layer connection can impact the frequency response of a system. In this figure the topology under consideration had a total system trace length of 22 in. and include two connectors. The trace in the topology resided either on the top or bottom layer.
Figure 4 shows the eye diagram after the equalizer for the topology where the trace resides on the bottom layer. This channel has a minimal degradation. It is clear that the ripple in the channel is not very large, hence the increased jitter in the eye is minimal.
Figure 5 shows the eye diagram for the same system conditions as Figure 4, but for a top-layer trace.
The above configurations illustrate that as the channel deviates from a trace-only characteristic, the eye becomes more difficult to recover, even with equalization techniques. The same is also true for copper cable. Figure 6 shows the eye diagram for a 10-Gbps data stream over 6 meters of 26 AWG differential cable and commonly used connector at either end.
Final Observations
As data rates have increased, the interconnect has evolved to the stage where active compensation for channel characteristics is required. Attention to passive channel characteristics is still required in order to allow the active techniques to work to their maximum effect. S-parameter criteria to evaluate the suitability of a channel for active equalization have been provided.
Above, we focused on the underlying copper structure or channel, but ultimately, it is the throughput of the system interconnect, which includes the driver and receiver devices, that needs to be maximized. Until recently, the main focus on driver devices was minimizing output impedance to maximize its drive capability. Conversely, the main focus on receiver devices was impedance matching to prevent reflections.
Drivers and receivers are now taking a more active role in maximizing a given interconnect's throughput capacity and reach. The use of techniques such as adaptive equalization, pre-emphasis, multi-level signaling, forward-error correction (FEC), and combinations thereof can help increase the throughput and/or reach through an appropriately designed channel. Thus, the system interconnect has evolved from a passive channel to an active channel, where the devices compensate for the channel or condition the data for the intended channel.
Receive-end adaptive equalization has been shown to be an effective method to overcome the ISI caused by the frequency dependent losses of a transmission media. In this case, amplification of the high-frequency content of the received signal is performed to recover the signal. This has the unfortunate effect of equally increasing high-frequency noise present in the system. This may require connector noise budgets tighter than what is traditionally considered conservative. Note that the increase of high frequency noise would also be characteristic of other active channel compensation systems such as pre-emphasis.
One can also see how the use of active techniques will therefore force the designer to further consider each aspect of the channel. Ultimately, it is important that the designer understand the synergistic performance between the compensation method employed by the chosen device(s) and the overall channel.
It has been demonstrated that the synergy of good passive design and active equalization can successfully transmit 10 Gbps serial data over copper, both traces and cable. This can be seen as key to increasing the capacity of copper backplanes, as well as enabling 10-Gbps small-form-factor (SFF) modules, both optical and copper. This also enables data rates greater than 10 Gbps by using multiple lanes of data, similar to the Ethernet XAUI interconnect.
Above, we focused on the underlying copper structure or channel, but ultimately, it is the throughput of the system interconnect, which includes the driver and receiver devices, that needs to be maximized.
We believe that the equalization method presented here combined with the optimization of the underlying copper structure will be able to further increase the data rates achieved over a single serial line.
References
- Fogg, Mike, "Evaluation of Maximum Usable Lengths for Cabled Copper Interconnects," http://www.tycoelectronics.com/products/simulation/files/papers/dc00mf.pdf.
- Morgan, Chad, and Helster, Dave, "The Impact of PWB Construction on High-Speed Signals," http://www.tycoelectronics.com/products/simulation/files/papers/dc99cmdh.pdf.
- Morgan, Chad, and Helster, Dave,"New Printed-Wiring-Board Materials Guard Against Garbled Gigabits", EDN, 11/99. Rothermel, Brent, Helster, Dave, and Sharf, Alex, "Practical Guidelines for Implementing 5 Gbps in Copper Today, and the Roadmap to 10 Gbps," http://www.tycoelectronics.com/products/simulation/files/papers/dc00brdh.pdf.
- Shakiba, Mohammad Hossein, "A 2.5 Gb/s Adaptive Cable Equalizer," ISSCC '99, Paper 23.3.
Editor's Note: This article is based on a presentation made at the DesignCon 2002 show, which is produced by the International Engineering Consortium (IEC; www.iec.org)
About the Authors
Ken Lazaris-Brunner is a product definition specialist with Gennum Corporation in their Datacom Products Division. Ken holds a B.Sc (Honours, Physics) degree from Queen's University (Kingston, Canada) and a M.Eng from McMaster University (Hamilton, Ontario). He can be reached at ken_la@gennum.com.
Bharat Tailor is a senior product manager with Gennum Corporation's Datacom Products Division. He holds a B.Eng (Honours) degree from McGill University (Montreal) and an MBA degree from Wilfrid Laurier University (Waterloo). Bharat can be reached at bharat_t@gennum.com.
John D'Ambrosia is the manager of semiconductor relations for Tyco Electronics. John received a B.S. in Electrical Engineering Technology from Pennsylvania State University in 1989 and a Master's Degree in Engineering Management from the National Technology University in 1998. He can be reached at john.dambrosia@tycoelectronics.com.
Michael Fogg is a member of the technical staff in Tyco Electronics' Circuits and Design Group. Michael has an M.S. degree in Engineering and a BS degree in Electrical Engineering from Pennsylvania State University. He can be reached at mike.fogg@tycoelectronics.com.
|
|
|
|
|
