Multiband architecture for high-speed SerDes

Abstract
As the speed of serializer/deserializer (SerDes) increases (e.g., to 25 Gbps and above), the channel will cause more severe inter-symbol interference. Design of low-complexity transceivers for such high-speed SerDes faces many technical challenges. In this paper, we explore a multiband architecture for a 25 Gbps SerDes, where the channel in each sub-band is approximately frequency flat, eliminating need of an equalizer in the receiver. Since different bands experience different signal attenuations, the power level for each band can be adjusted accordingly to minimize the average transmission power. A multiband transceiver is designed, and analysis and simulation results for various choices of parameters (bands, modulations, etc.) are presented.

1. Introduction
The demand for narrower interfaces in 100G Ethernet drives the need for high speed serializer/deserializer (SerDes) running at 25Gbps or higher.

This article is from a class at DesignCon 2011. Click here for more information about the conference.

However, there are many technical challenges in developing such a high-speed SerDes whose power consumption must be within acceptable limits.1 At rates as high as 25Gbps, the transceiver must be simple, since advanced signal processing at such speed is extremely power hungry. How to efficiently use the available bandwidth with low hardware complexity is therefore very important. The right signaling format or modulation scheme enables one to achieve such a goal.

Various modulation schemes could be employed for SerDes. Non-return-to-zero (NRZ) signaling has been used in most of the existing high-speed SerDes designs, and is the signaling format used by OIF CEI-11 and IEEE 10G Base-KR. However, there is controversy on if NRZ is the right signaling format for SerDes at 25Gbps and beyond.

First, the realistic channels one will be likely encounter are so lossy that a NRZ based receiver either cannot equalize them or requires an unrealistically complex equalizer which cannot be implemented within an acceptable power budget. The frequency responses of several channels contributed to IEEE802.3AP are shown in Figure 1 . As one can see, they have loss much higher than 25dB at 12.5GHz. Although these are legacy channels designed for 10Gbps operation, they have relatively short trace lengths. Channels using new connectors designed for 25Gbps operation and low loss channel materials such as megtron 6 also have loss in the neighborhood of 35dB at 12.5GHz when the reach is 40 inches and longer. As a result, these channels are representative of the channels with long reach as well as the channels with shorter reach, but using lower cost (and therefore higher loss) channel materials.

One can also see from Figure 1 that as frequency increases, the attenuation to the transmitted signal increases nonlinearly. This indicates that the channel is highly frequency selective, which results in severe inter-symbol interference (ISI). At 25 Gbps, ISI could span more than ten symbols. Therefore, a complex equalizer at the receiver is essential to ensure a very low bit error rate (BER) (e.g., ≤10-15 ). Implementation of such a complex equalizer at 25G either is not feasible or incurs a huge power penalty.

Secondly, the circuit implementations for NRZ are very challenging. Design of circuits such as capture latches, analog-to-digital converter (ADC) and transmit driver running at such data rate is very challenging. Although by using parallel circuits each block could run at a lower rate (e.g., 12.5 G or 6.25 G), it comes at the cost of power consumption and circuit area.

Figure 1: Channel frequency response.


Click on image to enlarge.

There are many types of higher order modulation schemes that are also applicable to SerDes. Pulse amplitude modulation (PAM), duo-binary (DB), quadrature amplitude modulation (QAM), phase shift keying (PSK), and constant envelope modulation such as frequency shift keying (FSK) are a few examples. A drawback of amplitude modulation is that for a given maximum voltage limit, the minimum distance between symbols will be reduced as the modulation level increases. For example, for the same maximum voltage limit as NRZ, the minimum distance of PAM-4 is only one third of that of NRZ, resulting in a 9.5 dB of signal-to-noise ratio (SNR) penalty up front. Constant envelope modulation overcomes the voltage limit problem of amplitude modulations. However, in a single- band approach, it is not easy to accurately control the phase or frequencies at such a high speed as 25 Gbps (e.g., non-pulse-shaped QPSK would require the signal phases to be 90 degrees apart).

All the modulation schemes mentioned so far fall in the category of single-band approach. Because the channel is highly frequency selective in the band from 0 to 12.5GHz, all these schemes cannot avoid the use of a multi-tap DFE, which is very power hungry.

In this paper, we explore a multiband architecture for SerDes at 25Gbps or beyond. In this architecture, the frequency spectrum is divided into several smaller sub-bands, where parallel data streams use different sub-bands. This architecture has many advantages over the single-band approach:

  1. in each sub-band, the channel is approximately frequency flat, consequently no equalizer in the receiver is needed;
  2. the symbol rate in each sub-band is much lower, which makes the circuit design easier;
  3. and it may occupy a smaller overall bandwidth if higher order modulation schemes are used in the sub-bands.

The idea of multi-band communications has been around for many years. It has been used extensively in wireless for example. However, the data rates are much lower there. In recent years, multi-band architectures at data rates higher than 1Gbps have also been studied. For example, a16-carrier QAM-16 architecture has been studied in optical communications.6 Different multiband architectures. such as analog multitone (AMT) have also been proposed for high-speed SerDes 4, 5 AMT is the analog variation of the discrete multitone while it tries to eliminate the use of ADC. However, a multi-input, multi-output (MIMO) DFE is required.

2. Background
2.1. Channel model
We consider a backplane channel over which the system should achieve a target BER not higher than 10-15 . The S-parameters obtained from measurements are provided by Intel (available in the endnotes).2 The test frequency ranges from 50 MHz to 15 GHz with a step size of 10 MHz. We extrapolate the data from 0 to 50MHz, where linear phase and 0 dB attenuation at ƒ = 0Hz are assumed. The resulting channel impulse response is shown in Figure 2 .

Figure 2: Backplane channel impulse response obtained from S-parameters.


Click on image to enlarge.

If the data rate is 25 Gps, Figure 2 indicates that significant ISI lasts at least 12 symbol periods (about 0.48 ns) with NRZ signaling.

2.2. Multiband vs. single-band approaches
To assess how severe ISI is at different signaling rates, we simulate the received signal in the absence of additive white Gaussian noise (AWGN). The eye diagrams of NRZ signals with four different data rates passing through the channel are illustrated in Figure 3 . It is observed that when the data rate increases to 10 Gbps, it is nearly impossible to achieve the target BER without an equalizer.

Figure 3: Eye diagrams of NRZ signals at four different data rates.


Click on image to enlarge.

If the signal occupies a smaller portion of the spectrum, it is possible that an equalizer might not be needed. Also, as shown in Figure 1, the attenuation to the transmitted signal at low frequencies is much lower than at high frequencies. Thus, it would be desirable to have the flexibility to adjust the transmission power levels of the signal that occupies different portions of the spectrum.

This could be achieved with a multiband approach. Figure 4 shows a general functional diagram of a multiband transmitter. The serial data are first converted into several sequences of parallel data streams. Then each sequence of data is modulated, using different modulation schemes, and pulse shaped by a low-pass filter (LPF). Pulse shaping filtering could be done in either the digital or the analog domain. After that, each sequence of data is moved to the desired sub-band. A traditional mixer is shown in Figure 4 for moving different data streams to different sub-bands. If the data rates for each band are not very high, it is possible to use digital methods for this conversion. Finally, the signals from different sub-bands are added together and transmitted.

Figure 4: Illustration of the multiband transmitter.


Click on image to enlarge.

The receiver of the multiband architecture is illustrated in Figure 5 . The receiver operates reversely to the transmitter. Signals of all sub-bands are converted to the baseband with mixers and LPFs. After that, data in each sub-band are demodulated to recover the original bit sequence.

Figure 5: Illustration of the multiband receiver.


Click on image to enlarge.

In the next section, we will discuss the detail of the multiband architecture by using the example of a 25 Gbps SerDes transceiver.3. Optimization of a Multiband Architecture for 25 Gbps SerDes
3.1. Transmitter processing
Depending on how many sub-bands are used and the modulation scheme adopted (e.g., binary or higher-order QAM), the signal in each sub-band might still experience slight amount of ISI because of the non-flat channel frequency response in each sub-band. Given the frequency-attenuation profile of the backplane channel as shown in Figure 1, the error performance of the system will remain to be ISI dominated. If minimizing this slight ISI still turns to be necessary, a pre-emphasis unit for each sub-band could be employed at the transmitter. This unit approximately equalizes the channel gain in each sub-band. Since the bandwidth of each of the bands is not very large (e.g., not exceeding 4 GHz or so), design of such unit in the analog doma.in with low complexity is practical.

The transmitter model with 4 sub-bands for the 25 G SerDes is shown in Figure 6 , in which a pre-emphasis unit in the transmitter is assumed; the case without a pre-emphasis unit is a special case in which this unit has a constant gain at all frequencies in band.

Note that this specific configuration is only for illustration of the architecture; the number of sub-bands could be flexibly chosen to accommodate different settings. Indeed, in Section 4, we will evaluate the performance of two variations of this architecture: a system with four sub-bands but no pre-emphasis blocks, and a system with three sub-bands but with pre-emphasis blocks. Four sub-bands are considered because eye-diagram simulation results reveal that even when binary PAM is used in band 1 and QPSK is used in bands 2 to 4, the channel-induced ISI would be at a tolerable level; three sub-bands with a simple pre-emphasis unit are considered because the target BER performance can still be maintained without an equalizer at the receiver.

Figure 6: The 25G multiband transmitter model.


Click on image to enlarge.

Based on the channel model in Sections 1 and 2, we optimize modulation, bandwidth, guard-band, and power level for each sub-band. First, we design the pulse shape for the signals to be transmitted in each sub-band. Here, for baseband pulse shaping we adopt the raised-cosine filter (RCF), which is widely used in communication systems and standards. The impulse response of the RCF is expressed as:


Click on image to enlarge.

where T is the symbol period and β is the roll-off factor, which determines how much excess bandwidth is required over the minimum bandwidth (i.e., an ideal filter with a sinc impulse response) required to avoid ISI. The roll-off factors for each sub-band do not have to be identical.

To maintain zero-ISI, the desired pulse shape is the composite pulse in the receiver after receive baseband filtering and includes the effect of the channel. Since the channel will not have a perfectly flat frequency response for each sub-band, small amount of ISI exists.

To maximize the signal-to-noise ratio (SNR) at the receiver, it is common in practice to split an RCF as a pair of square-root-raised cosine filters (SRCFs), one at the transmitter and one at the receiver. The SRCFs at the transmitter and receiver are also known as the matched filters. The impulse response of an SRCF is expressed as:


Click on image to enlarge.

Note that for bandwidth not exceeding 3-4 GHz, implementation of very small passive analog filters that approximate the ideal frequency response of the SRCFs is practical.

As shown in Figure 6, one data stream that uses the lowest frequency band is transmitted using baseband signaling. As a result, for a transmitter with N bands, the transmitter only employs N -1 mixers.

In addition to approximately equalizing the channel gain over each sub-band, the pre-emphasis modules are also used to adjust the signal power in each sub-band. Since the channel behaves like a low pass filter, signals transmitted in a sub-band in the higher frequency region are attenuated more than those in the lower frequency area. To enhance the signal at high frequency area, we multiply the signal at each sub-band (except the baseband) with a coefficient Cpre_em . If f1 < f2 < f3, then the coefficients satisfy:


Click on image to enlarge.

3.2. Receiver processing
The receiver model is shown in Figure 7 .

Figure 7: The multiband receiver model.


Click on image to enlarge.

It is almost the same as Figure 5, except that the LPFs are replaced by SRCFs.

It is critical to examine how ISI could be minimized in the sub-bands so that the receiver does not need an equalizer. Thus extensive experiments (simulation) are conducted to determine the near-optimal system parameters. The general specifications and guidelines are summarized as follows:

  1. The minimum overall aggregated data rate is 25 Gbps;
  2. The target BER is 10-15 . However, since using simulation to obtain such a low BER takes too long, some semi-analytic or qualitative approaches will be used to assess the approximate BER performance.
  3. The number of sub-channels should be minimized to reduce the complexity;
  4. The roll-off factor of the SRCFs should be as small as possible to minimize the total bandwidth, but too small a roll-off factor might lead to an unacceptable level of ISI.

4. Simulation results
4.1. Architecture with four sub-bands and without pre-emphasis
4.1.1. System with binary modulation

The first architecture only employs binary modulation schemes for each dimension of the signal, in order to reduce complexity; that is, NRZ is employed in band #1 and QPSK is employed for all other bands. One of the advantages of this architecture is that backward compatibility can be maintained by turning on only the baseband to receive the signal from the legacy transmitter. Furthermore, the circuit implementation is closest to the existing NRZ one at lower data rates since for I or Q phase of the QPSK, a NRZ signal is used. Simulation results show that without a pre-emphasis block, there should be as least four sub-bands to achieve a virtually error-free transmission. Parameters of each sub-band are listed in Table 1 .3

Table 1: Parameters of the multiband transceiver with binary modulation (per real dimension).


Click on image to enlarge.

The overall data rate of this architecture is 25.3 Gbps.

If there were no ISI and potential bit errors were only caused by AWGN, the bits transmitted in the worst-performance band, that is, the band with a center frequency of 13.3 GHz, would virtually be error free even with a low transmission power. This is briefly analyzed next. Let us assuming the following system parameters:

  • Transmission power: Pt = 20mw, which roughly corresponds to an average signal voltage of 1 volt across a 50Ω load;
  • Channel loss: L = 60 dB, which corresponds to the worst case – band #4. Of course, if necessary, the transmission power could be increased;
  • Receiver noise figure: Nf = 6dB.

The data rate and bandwidth are given in Table 1. The received SNR (which also equals Eb /N0 for this case) can be calculated to be 32.49 dB, sufficient to achieve virtually error free transmission in AWGN. This indicates that if ISI is small, the transmitter might not need a power amplifier to generate a signal at such power/voltage level. Note that in this calculation, since an RCF pulse shape is adopted, the noise bandwidth is 3.55 GHz, rather than 4.97 GHz.

The signal spectrum at the receiver (after the channel) is shown in Figure 8 .

Figure 8: Power spectrum of the received signal with binary modulation (per real dimension) schemes.


Click on image to enlarge.

From Figure 8, we observe that the overall required bandwidth of this architecture is 15.85 GHz, which is larger than 12.5 GHz for a single band NRZ SerDes. However, in each sub-band, the symbol rate is much lower, which makes it easier to implement.

In order to assess the ISI level, the scatter plot of the received symbols (1+i , 1-i , -1+i , -1-i ) in the absence of AWGN for the signal transmitted in band #1 is shown in Figure 9 . The effects of the channel and pulse shaping filters are included in this plot. Band #2 is chosen because it has two adjacent sub-bands that interfere with it, which represents the worst case in terms of ISI and adjacent-band interference; other bands have similar or slightly better scatter plots.

Figure 9: Scatter plot of received QPSK symbols in the band #2.


Click on image to enlarge.

It is observed that ISI reduces the minimum distance between symbols by 75%, resulting in a reduction of the effective Eb /N0 of 12 dB. From the simple analysis above, given a low transmit power, this reduced minimum distance between symbols is still sufficient to maintain a very low BER when AWGN is considered.

It is worth to mention that a mixer with a central frequency of 13.3 GHz is technically feasible but not preferable. Also mixers consume power. Data transmitted at such a high frequency may experience other distortions. In the next subsection, we will try to assess the possibility to reduce the overall required bandwidth by employing higher order modulation schemes in the architecture.4.1.2. System with higher order modulation
After optimization of the architecture through experiments, the multiband architecture employing higher order modulation schemes is summarized in Table 2 .3

Table 2: Parameters of the multiband transceiver with higher order modulation schemes ,


Click on image to enlarge.

The overall bandwidth of this architecture is 11.9 GHz, which is slightly smaller than 12.5 GHz, but much less than the architecture employing binary modulation scheme. As for the case of binary modulation in the previous section, the signal spectrum at the receiver (after channel) of this architecture is shown in Figure 10 . Compared with Figure 8, the power spectrum is more compact. An analysis of the channel reveals that within the frequency range from 7 GHz to 11 GHz, the channel is flatter and smother than those at other frequencies. This indicates that transmitting more data in this frequency range is possible. Therefore, the bandwidth of the 3rd and the 4th sub-bands are made wider than that of the 2nd sub-band.

Figure 10: Power spectrum of the received signal with higher order modulation schemes.


Click on image to enlarge.

Unfortunately this architecture also has a drawback: we need to use 4-PAM scheme for the baseband and 16-QAM for other bands to maintain a good BER performance as well as high bandwidth efficiency. Again, in order to assess the impact of ISI and adjacent-band interference, the received symbol scatter plot for band #2 is shown in Figure 11 .

Figure 11: Scatter plot of received 16-QAM symbols in band #2.


Click on image to enlarge.

4.2. Architecture with three sub-bands and pre-emphasis
The benefit and feasibility of implementing pre-emphasis modules for each sub-band is briefly discussed in Section 3.1. Simulation experiments reveal that with a simple pre-emphasis unit for each sub-band to approximately flatten the channel frequency response, three sub-bands are sufficient to achieve 25 Gbps SerDes without an equalizer in the receiver. For this case, we only present the results with the same modulation as that chosen in Section 4.1.2. Other system parameters are given in Table 3 .

Table 3: Parameters of the multiband transceiver with higher order modulation schemes (with transmitter pre-emphasis modules for each sub-band).


Click on image to enlarge.

The signal spectrum at the receiver of this architecture is illustrated in Figure 12 , where for the purpose of assessing the sub-band spectrum, a gain is applied to the transmitted signals.

Figure 12: Power spectrum of the received signal with higher order modulation schemes (with transmitter pre-emphasis modules).


Click on image to enlarge.

It is observed that due to the pre-emphasis unit, the spectrum in each sub-band is much flatter than the cases in Section 4.1. Therefore, it is expected that ISI will be minimal with this architecture. Also, since the spectrum allows a large guard-band between adjacent sub-bands, adjacent band interference will be negligible as well. To assess the severity of ISI and adjacent band interference, again the scatter plot of the received symbols in the second band is shown in Figure 13 .

Figure 13: Scatter plot of received 16-QAM symbols in the band #2 (with a pre-emphasis module at the transmitter) .


Click on image to enlarge.

Compared with Figure 11 (the case without a pre-emphasis unit), Figure 13 shows that ISI level is significantly reduced by the pre-emphasis unit.5. Discussion
For SerDes with rates of 25 Gbps and beyond, multiband transmission could be an attractive low-computational requirement, low-power alternative. The major benefits are:

  1.  a computational-demanding equalizer for single-band approaches is not needed,
  2. processing in each sub-band is significantly reduced when compared with the single-band approach, and
  3. transmission power level can be adjusted to the specific channel characteristics of each sub-band to lower the required average transmission power levels.

However, there are added complexities in this architecture as well:

  1. mixers and oscillators are in general needed and
  2. tracking carrier frequency and phase is required at the receiver.

The current observation is that when speed increases to 25 G and beyond, the benefits of a multiband approach prevail the associated drawbacks, and is thus preferred over a single-band approach. To ultimately realize the benefits of a multiband system, circuit design must also be optimized. For example, can the transmit and receive pulse shaping filters be realized in the analog domain with sufficient accuracy? Can the mixer, oscillators, and carrier tracking (frequency and phase) for each sub-band be implemented with low complexity and low power consumption? These should be studied further in hardware experiments.

6. Conclusion
In this paper, we have explored the possibility of using a multiband architecture for high speed SerDes. A number of cases have been considered: three and four sub-band architectures, with binary modulation (per real dimension) and with higher order modulations (e.g., 16-QAM), and with/without transmitter pre-emphasis modules. Simulation results have shown that with an optimized multiband architecture and a low transmission power, SerDes could achieve virtually error free communications at 25Gbps over channels which are too challenging for SerDes employing legacy signaling formats.

Christian Weber received the Dipl.-Ing.(FH) degree in communication engineering from University of Applied Sciences, Offenburg, Germany, in 2009. From Feb. 2009 to Sep. 2009, he was with Oregon State University, Corvallis, USA, working on a multiband architecture for high-speed SerDes. His research interests are radio frequency, filter banks and spectrum sensing in cognitive radio.

Jinjin He received the B.S. and M.S. degrees in electrical engineering from Fudan University, Shanghai, China, in 2003 and 2006, respectively, and the Ph.D. degree in electrical engineering from Oregon State University, Corvallis, in 2010. She has been with Marvell since July 2010, working on flash memory controller chip. Her technical interests are low-power/high-speed VLSI design for digital signal processing and error correcting techniques in communication systems and storage systems.

Charlie Zhong has a B.S. degree from Tsinghua University, China, a M.S. degree from NJIT and a Ph.D. degree from UC Berkeley, all in electrical engineering. Since January 1995, he has worked in the AT&T Bell Labs, Berkeley Concept Research Center, ST Microelectronics, Avnera, W5 networks and MediaPHY before joining LSI, where he is currently a principal system engineer. His interests include high speed SerDes and wireless networks. He has two awards from the Bell Labs in 1998 for extraordinary achievements on the CDMA program, and the Management of Technology certificate from the Haas School of Business in 2003. He has also received CSG value recognition award from LSI for innovation on high speed SerDes in 2006. In addition, he currently holds fifteen patents in wireless communications and high speed SerDes.

Huaping Liu received the B.S. and M.S. degrees in electrical engineering from Nanjing University of Posts and Telecommunications, Nanjing, China, in 1987 and 1990, respectively, and the Ph.D. degree in electrical engineering from New Jersey Institute of Technology, Newark, in 1997. From July 1997 to August 2001, he was with Lucent Technologies, Whippany, NJ. Since September 2001, he has been with the School of Electrical Engineering and Computer Science, Oregon State University, Corvallis, where he is currently an Associate Professor. His research interests include ultrawideband systems, multiple-input multiple-output antenna systems, channel coding, and modulation and detection techniques for multiuser communications. He is currently an Associate Editor for the IEEE Transactions on Vehicular Technology and IEEE Communications Letters, and an Editor for the Journal of Communications and Networks.

Endnotes:

  1. C. Zhong, C. Y. Liu, W. Jin, A. Malipatil, G. Tang, and F. Y. Zhong, “Design considerations in high speed SerDes (25 Gbps),” in Proc. DesignCon 200 8.
  2. Available at http://ieee802.org:80/3/ap/public/channel_model/index.html
  3. C. Weber, “Multiband Architecture for a 25 Gbps High Speed SerDes on a Backplane Channel: Analysis, Simulation and Specification,” MS thesis, Offenburg University of applied Sciences, Germany and Oregon State University, USA, Sep. 2009.
  4. A. Amirkhany, A. Abbasfar, V. Stojanovic´, and M. A. Horowitz, “Analog multi-tone signaling for high-speed backplane electrical links,” in Proc. IEEE GlobeCom'06 , , Nov. 2006, pp. 1–6.
  5. W. T. Beyene and A. Amirkhany, “Controlled intersymbol interference design techniques of conventional interconnect systems for data rates beyond 20 Gbps,” IEEE Trans. Advanced Packaging , vol. 31, no. 4, pp. 731-740, Nov. 2008.
  6. J. Chen, et al,. “A multi-carrier QAM transceiver for ultra-wideband optical communication,” IEEE J. Solid State Circuits , vol. 41, no. 8, pp. 1876-1893, Aug. 2006.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.