Feasibility of 25 Gb/s Serial Transmission Over Copper Cable Assemblies

Transcription

1 DesignCon Feasibility of 5 Gb/s Serial Transmission Over Copper Cable Assemblies Vittal Balasubramanian, FCI USA, LLC [email protected] Stephen B. Smith, FCI USA, LLC [email protected]

2 Abstract: In the quest to transmit higher data rates over high-speed cable assemblies, various standards organizations such as the IEEE 8.3ba Committee are proposing extensions to existing specifications. One possible path to achieving transmission at Gb/s would be to move to a x 5 Gb/s implementation. To accomplish this, it is first necessary to define performance requirements for each of the components of the cable assembly. A new strategy to accomplish this is to attack the problem from a bottom-up approach whereby a theoretically perfect cable assembly is first characterized to establish the feasibility of transmitting signals at a data rate of 5 Gb/s. Subsequently parameters of real components are inserted into the assembly and the resultant performance degradation is observed. The analysis yields a set of requirements for each of the individual components and a determination of the minimum signal conditioning requirements. Author s biography Vittal Balasubramanian received his B.S. in Electrical Engineering from Delhi College of Engineering, University of Delhi in and his M. Eng. in Electrical Engineering from Penn State University in 5. He was awarded the Doris Hughes Memorial Award at Penn State in recognition of his outstanding academic achievements and contribution to the college community. He has been working at FCI USA, Inc. since Jan 5 where his responsibilities include the design and analysis of high-speed connectors and cable assemblies. He also worked for Sapient Corporation Pvt. Ltd. from August to December as a technology associate and helped with software consulting. He is a member of the Institute of Electrical and Electronics Engineers since 998. Stephen B. Smith s current responsibilities at FCI include primarily customer support in the application of high-speed connectors by working very closely with Marketing and Sales. While previously working in the development arena, he performed the electrical design work on FCI s popular AirMax VS Backplane Connector. Prior to coming to FCI in, Smith worked at AMP Incorporated (now Tyco Electronics) for years as a development engineer spending most of that time in the electromagnetics research group where he developed methods of modeling and simulating interconnection systems on projects spanning the frequency spectrum from power-frequency high-current utility connectors to high-speed and r.f. interconnects. Prior to that, he worked in acoustical research at Masland Industries (now Lear Corporation.) He has taught various courses at each of his jobs, and in the last couple of years, he has presented papers at various conferences including the IEEE Holm Conference. Smith has a B.S. in physics and an M.S.E.E., both from Penn State University.

3 . Introduction Today s high speed cable IO s are designed to work with a throughput varying from Gb/s, Gb/s, Gb/s, and even to Gb/s over lengths of up to m of copper cable (SFP+, QSFP, IEEE 8.3ba, CXP). These throughputs are realized by implementing a PHY with,, or parallel lanes respectively, each working at Gb/s. One logical progression which will likely be considered by some standards organizations would be to move to a x 5 Gb/s implementation which would reduce the signal density requirements and routing complexity. However, such an implementation would not be straightforward, as it would be fraught with many signal-integrity and other challenges that would need to be solved. Such challenges would include the development and selection of transceiver, cable, cable termination, and connector technologies. When attempting to increase the speed of each individual serial channel to 5 Gb/s, the performance of the cable assembly poses a significant bottleneck, and therefore it is necessary to characterize its performance. This paper addresses the SI challenges inherent in such high-speed transmission and proposes requirements for each of the different components in a passive 5 Gb/s copper cable link. The authors propose a new strategy for deriving component requirements. Rather than using the traditional method of performing simulations with existing components to determine whether a link with these components would work, a structured bottom-up approach is proposed that begins by using theoretical models of ideal components in order to establish the feasibility of transmission at 5 Gb/s. Incrementally, models of real components are introduced into the overall cable assembly model so that the effects of each incremental addition can be observed. Seeing the effects of adding each individual real component model makes it possible to determine the actual component requirements to achieve x 5 Gb/s signal transmission over a copper cable.. Feasibility of transmission at a data rate of 5 Gb/s through a cable assembly While there are no published industry standard documents defining the requirements for the performance of a 5 Gb/s cable assembly, several committees (IEEE 8.3ap, SAS G, FiberChannel, OIF CEI 5G-LR, and OIF CEI 8G-SR) have proposed criteria for characterizing channels comprised of backplanes that are designed to operate at data rates between and 5 Gb/s. Additionally, the IEEE 8.3ba Committee is working to establish performance criteria for cable assemblies that will operate at a speed of Gb/s. Since the specific topic of this study is cable assemblies, we will first consider the IEEE P8.3ba requirements and will extrapolate the performance limits therein to provide new criteria which are more relevant to the transmission of speeds up to 5 Gb/s. Following this, we will use the OIF CEI 5G-LR standard which provides reasonable performance criteria for the speeds of interest despite being exclusively defined for backplane applications. Only cable assemblies with loss characteristics falling within the limit range defined by this standard will be considered, hence the length of the cable assembly will be appropriately limited. 3

4 . Extrapolated IEEE spec The IEEE P8.3ba specification, which pertains only to a serial data rate of Gb/s per lane, places limits on four parameters: Insertion Loss, Insertion Loss Deviation (ILD), Return Loss (RL), and Integrated Crosstalk Noise (ICN). In order to extrapolate the IEEE P8.3ba spec from Gb/s to 5 Gb/s, all high end frequencies in the spec limits are multiplied by a factor of.5. Figure shows the original and the extrapolated limits intended to extend the frequency range of interest from the 7.5 or GHz in the spec all the way up to frequencies more commensurate with the target data rate of 5 Gb/s. Figure a shows that the Insertion Loss limit has been extended up to 8.75 GHz. Requiring the same performance at 8.75 GHz that was previously required at 7.5 GHz has the effect of significantly tightening the performance requirements over the entire frequency range including the lower frequencies below 7.5 GHz. The same is true for the extrapolation and tightening of the ILD limit shown in Figure b. In the IEEE P8.3ba spec, the insertion loss deviation (ILD) must be between ±.7 db at DC and between ±. db at 7.5 GHz. In the 5 Gb/s spec the ILD must still be between ±.7 db at DC, but it must now be between ±. db at 8.75 GHz (see figure b). Insertion Loss [db] - -5 Insertion Loss Gb/s Spec 5 Gb/s Spec Insertion Loss Deviation [db] Insertion Loss Deviation Gb/s Spec 5 Gb/s Spec Return Loss Gb/s Spec 5 Gb/s Spec INTEGRATED CROSS-TALK NOISE ICN Limit Return Loss [db] ICN [mv, RMS] IL at.9 GHz [db] Figure : Gb/s IEEE P8.3ba spec and its extrapolation to 5 Gb/s: Shown are Insertion Loss/attenuation specifications, Insertion Loss Deviation specifications, Return Loss specifications, and Integrated Crosstalk Noise specification. The Return Loss limit, shown in Figure c, is extended to 5 GHz from the previous GHz required in the IEEE P8.3ba spec. Finally, the requirement for ICN is tightened by calculating it at a frequency of.8965 GHz rather than the GHz listed in the IEEE P8.3ba spec. Additionally, the computations of multiple disturber near-end

5 and far-end crosstalk are performed to a frequency of 5 GHz as opposed to the original GHz in the IEEE P8.3ba spec. The limits for ICN, however, remain the same for the higher data rate, assuming similar loss and crosstalk budgets are needed.. OIF CEI 5G-LR spec Although as noted above, the OIF CEI 5G-LR specification pertains to channels consisting of backplanes, it will be applied here in the context of cable assemblies by virtue of the fact that it is the only existing specification presently under consideration that applies to a data rate up to 5 Gb/s. The limits for IL, ILD, RL and ICN are shown in Figure. The extrapolated IEEE spec is more stringent than the OIF spec for the IL and the ILD. In the extrapolated IEEE spec, the ILD must be better than ±.7 db at DC and better than ±. db at 8.75 GHz. Above 8.75 GHz, there is no spec for the ILD. In the OIF spec, by contrast, the ILD limit reaches its maximum of ± db at 6 GHz and must remain better than ± db all the way up to 5 GHz. INSERTION LOSS - ATTENUATION 6 INSERTION LOSS DEVIATION Insertion loss [db] Return loss [db] CEI-5G LR Limits Extrapolated IEEE 5 Gb/s Spec RETURN LOSS CEI-5G LR Limits Extrapolated IEEE 5 Gb/s Spec Insertion loss deviation [db] ICN [mv, RMS] - - CEI-5G LR Limits Extraploated IEEE 5 Gb/s Spec INTEGRATED CROSS-TALK NOISE CEI-5G LR Limits 5 5 IL at.9 GHz [db] Figure : Comparison of OIF CEI 5G-LR spec to the extrapolated IEEE spec: Insertion Loss/attenuation specification, Insertion Loss Deviation specification, Return Loss specification, and Integrated Crosstalk Noise specification. Note that the ICN limit in the OIF specification is the same as that in the extrapolated IEEE specification. 5

6 As of late 9, there are no transceivers commercially available on the market that can transmit at a data rate of 5 Gb/s. Therefore it is necessary to evaluate the performance of the cable assembly strictly through simulation and analysis rather than measurement. The proposed bottom-up analysis approach starts with a perfect link consisting of an ideal cable, ideal component boards (paddle cards), and ideal connectors with perfect footprints, all with no losses. One by one, the ideal components are replaced by real (non-perfect) components including cables and PCBs with increasing loss, impedance mismatches such as those caused by footprints, and crosstalk arising in the footprints and connectors. The performance degradation caused by each of the incrementallyintroduced imperfections will be compared against the two passive channel specifications. Additionally, the channel models will be used in Bit-Error Rate (BER) simulations at 5 Gb/s. The BER simulations will be run in combination with various signal-conditioning schemes such as FFE and DFE. Although the overall link performance can be affected by interdependencies between the components, the incremental addition of real component models should yield a set of performance requirements for each of the individual components. 3 Transceiver parameters Driver and receiver parasitics are taken into account by adding a capacitance to ground at both ends of each trace. All traces are terminated by a 5 Ohms resistor to ground in parallel with a.5 pf chip capacitance. Figure 3 shows the input RL of the transmitter and the receiver compared with the OIF CEI-5G-LR spec. Using the above mentioned value for the chip capacitance the input RL meets the OIF CEI-5G-LR spec. The driver transmits a NRZ bit stream with a bit rate of 5 Gb/s and a -9 % rise time of ps. Figure a shows the BER eye and figure b shows the BTC for a perfect lossless channel, but taking into account transceiver parasitics and driver jitter. The BTC has a width of about.7 UI for a BER of e, which is in line with both the extrapolated IEEE spec and the OIF CEI-5G-LR spec. Some kind of signal conditioning will be required to open the BER eyes at 5 Gb/s. In our link simulations the driver will use a baud-spaced pre-emphasis (PE) filter with up to 5 taps and the receiver will use a baud-spaced Decision-Feedback Equalizer (DFE) with up to 7 taps. RETURN LOSS Return loss [db] C =.5 pf CEI-5G-LR TX RL Spec Figure 3: Transmitter and receiver input RL vs. OIF CEI-5G-LR spec. 6

7 BER Sampling point [UI] Figure : BER eye and BTC in case of a perfect lossless channel.. Procedure for modeling the cable assemblies The scope of this effort includes modeling a cable assembly from the board-side connector on one end to that on the other end. Board traces between the transceivers and the board-side connectors which have loss characteristics in compliance with the fixture requirements of the IEEE 8.3ba spec have been included in the analysis.. Description of the cable assembly to be modeled As the IEEE 8.3ba Committee is presently using Quad Small Form Factor Pluggable (QSFP) cable assemblies in their evaluations, such assemblies will be considered here as well. Specifically, we will analyze QSFP cable assemblies of length 3, 5, 7, and m and which utilize raw cable of the appropriate wire gauges ranging from 3 to 6 AWG commensurate with what is required to achieve performance at the desired lengths. When beginning work on a new industry standard, committees typically start with an existing standard and make the limits from that standard more stringent to fit the new application. The next logical step is then to evaluate existing components against these more stringent requirements with the hope that they will meet them. When making the jump from Gb/s to 5 Gb/s, such a hope is perhaps overly optimistic. As mentioned earlier, rather than following this more traditional top-down approach of testing an existing cable assembly against new requirements, we will utilize a systematic bottom-up procedure in which we begin by analyzing a theoretically perfect cable assembly and then incrementally adding in real components in the order of raw cable, paddle cards, soldered termination of the cable to pads on the paddle cards, and finally board-side I/O connectors in order to observe the incremental channel degradation that occurs. At each step along the way, we will compare the performance against both the extrapolated IEEE P8.3ba spec and the OIF CEI 5G-LR spec (hereafter referred to as OIF). By breaking the problem down into the characterization of individual constituent components and adding them together one at a time, this approach allows us to determine some performance requirements for individual components of the overall cable assembly. 7

8 . Description of and results for each of the models A theoretically-perfect cable assembly would add nothing to the transmission channel other than simple delay without any distortion or loss in the signal. Hence, the output signal would be identical to the input signal and there is no need to evaluate this initial model... Addition of raw cable to the theoretically-perfect cable assembly model The first practical model to evaluate consists of that for a raw cable. The cables used in this exercise are ParaLink 7 and ParaLink -3 from LEONI AG. They consist of 8 differential pairs, each of ohms. The wire sizes are 3AWG and 6AWG, and the cable is standard Flame Retardant Non-Corrosive (FRNC). We consider lengths of 3, 5, 7 and m. The ParaLink 7 and ParaLink -3 cables were selected because they purport to have fairly smooth attenuation curves up to frequencies of about 7 and 3 GHz respectively, thus making them appropriate for usage at the high data rates relevant to this study. LEONI was gracious enough to provide Touchstone models of these particular cables in a few lengths. These models were used to create per-unit-length models for each type and gauge of cable. Subsequently, these per-unit-length models were used to create raw cable models of the various lengths shown in Figure 5. Figure 5 shows plots of the fitted Insertion Losses of these cables for the four lengths of interest compared against the extrapolated IL limit. (Note that in both the extrapolated IEEE spec and the OIF spec, only the curve fit of the IL is required to meet the limits, and not the actual IL data itself.) It is not necessary to show plots of ILD, RL, or ICN since there are not yet any impedance mismatches in these models nor is there any crosstalk. Insertion Loss [db] - Insertion Loss Gb/s Spec 5 Gb/s Spec m 3AWG ParaLink7 3m 3AWG ParaLink7 5m 3AWG ParaLink7 7m 3AWG ParaLink7 Insertion Loss [db] - Insertion Loss Gb/s Spec 5 Gb/s Spec m 6AWG ParaLink-3 ParaLink-3 ParaLink-3 7m 6AWG ParaLink

9 Insertion Loss [db] - Insertion Loss Gb/s Spec 5 Gb/s Spec m 3AWG ParaLink-3 3m 3AWG ParaLink-3 5m 3AWG ParaLink-3 7m 3AWG ParaLink Figure 5: Gb/s IEEE P8.3ba spec and its extrapolation to 5 Gb/s: Shown are Fitted Insertion Loss curves for 3AWG ParaLink 7, fitted Insertion Loss curves for 6AWG ParaLink -3 and fitted Insertion Losses for 3AWG ParaLink -3. Figure 5 shows that the following raw cables and lengths meet the extrapolated IEEE P8.3ba specification: i) 3m 3AWG ParaLink -3 ii) ParaLink -3 The ParaLink -3 cable marginally fails the IL requirement below.5 GHz but for most of the frequency range, meets it. Hence these three cables will be used in all further analysis. Based on the observed trend in Figure 5, it is possible that AWG cables would be able to achieve a longer reach, perhaps up to 7m or more. However, in the absence of models, this cannot be proven. Figure 5 demonstrates that the cable attenuation presents a significant limiting factor to the overall channel performance. Cable lengths longer than 5 m would not likely successfully transmit a signal at a serial data rate of 5 Gb/s. Hence, in the same way that the IEEE P8.3ba specification lists 7m as the cable assembly length for its longest reach at Gb/s, and the OIF specification limits the length of a backplane channel to approximately 7 cm at 5Gb/s, a future specification for cable assemblies transmitting a signal at a serial data rate of 5 Gb/s might have to limit the reach to 5m. Turning now to the OIF specification, Figure 6 shows plots of IL, fitted attenuation, and ILD for a 3m length of 3AWG ParaLink -3 cable compared against the appropriate limits. While the actual IL meets the OIF limits (Figure 6a,) the linear curve fit as defined by the OIF specification does not. Since the OIF specification applies to backplane channels, the expected shape of the IL curve at low frequency differs from that of cable assemblies. Conduction losses at low frequencies for cable assemblies are more prominent than they are for backplanes. This can be attributed to the generally lower value of dielectric loss tangent applicable to cable assemblies. Unlike backplanes, the low-frequency skin effect losses, which are proportional to f, are more prominent in cable assemblies. Hence, the attenuation curve fit as calculated per the OIF specification is not representative of cable assemblies. Likewise, Figure 6b shows the inappropriateness of determining ILD by using the attenuation curve fit as defined by the OIF specification, for cable assemblies. To remedy this inconsistency, this paper uses the 9

10 Insertion loss [db] Insertion loss [db] attenuation curve fit as defined by the extrapolated IEEE P8.3ba specification even when comparing against the OIF limits. Following this approach for the two raw cables that passed the extrapolated IEEE P8.3ba specification as well as the third cable that nearly passed, demonstrates that all three pass the OIF specification as shown in Figure INSERTION LOSS - ATTENUATION -3 CEI 5G-LR Limits -35 Fitted Attenuation 3m 3AWG ParaLink-3 Cable Insertion loss deviation [db] INSERTION LOSS DEVIATION CEI 5G-LR Limits ILD - 3m 3AWG ParaLink Figure 6: 3m length of 3AWG ParaLink -3 cable compared to the OIF CEI 5G-LR Specification: Insertion Loss - fitted attenuation, Insertion Loss Deviation. INSERTION LOSS - ATTENUATION -3 CEI 5G-LR Limits ParaLink-3-35 ParaLink-3 3m 3AWG ParaLink Insertion loss deviation [db] INSERTION LOSS DEVIATION CEI 5G-LR Limits ParaLink-3 ParaLink-3 3m 3AWG ParaLink Figure 7: Comparison to OIF CEI 5G-LR specification for various cables using extrapolated IEEE P8.3ba curve fit for fitted attenuation, and Insertion Loss Deviation. As seen in Figure 7, despite the failure at low frequencies of the 5m 6 AWG raw cable model compared against the extrapolated IEEE IL specification, all three models easily pass the OIF IL requirement... Addition of paddle cards After characterizing models of raw cables, the next step was to add models of the traces on the paddle cards. These models were generated by using a quasi-static -D analysis tool that is proprietary to FCI. To use this tool, the user defines the -D geometry of the paddle card stack up and the traces as well as the appropriate material properties for the pc board laminates. The solver computes the impedance and propagation delays. Knowing the loss permits the tool to effectively extrude the geometry into the third dimension to obtain the S-parameters. In the real cable assembly being modeled, the lossy traces on the paddle cards have a length of 6 mm on each paddle card.

11 The board material used for the paddle cards was FR, with a dielectric constant of and a loss tangent of.3. Microstrip traces were used on the paddle cards. Cascading the Touchstone (S-parameter) model of the paddle cards at each end of the model for the raw cable results in a cable assembly model that performs as shown in Figures 8 and 9 compared against the extrapolated IEEE and OIF specs respectively. The inclusion of the trace models presents at least one possible impedance mismatch at each end of the cable, but neither the ILD nor the RL are unduly disrupted. Insertion Loss [db] Insertion Loss Gb/s Spec 5 Gb/s Spec 3m 3AWG Insertion Loss Deviation [db] Insertion Loss Deviation Gb/s Spec 5 Gb/s Spec 3m 3AWG Return Loss Return Loss [db] Gb/s Spec 5 Gb/s Spec 3m 3AWG Figure 8: Performance of various lengths of ParaLink -3 cables with paddle card attached compared to Gb/s IEEE P8.3ba spec and its extrapolation to 5 Gb/s. Shown are IL - fitted attenuation, Insertion Loss Deviation, and Return Loss. Insertion loss [db] - -5 INSERTION LOSS - ATTENUATION -3 CEI 5G-LR Limits -35 3m 3AWG Figure 9: Comparison to OIF CEI 5G-LR specification for various lengths of ParaLink -3 cables with paddle card attached using extrapolated IEEE P8.3ba curve fit for fitted attenuation.

12 ..3 Addition of cable wire terminations To represent the soldered terminations of the cable wires to the pads on the paddle cards, a Touchstone model of the termination was obtained by analyzing a solid model using CST MicroWave Studio. Figure shows the solid model. Basically the model consists of the stripped end of the raw cable such that the bare wires protrude and are attached to pads on the paddle card. Figure : 3-D model of a cable wire to paddle card termination. After the addition of this component model at each end between those of the raw cable and paddle cards, all parameters were plotted against the two specs and the results shown in Figures and. The presence of the cable terminations adds more impedance discontinuities. Additionally, there is now crosstalk in the overall model, and hence, it is now appropriate to plot ICN. Insertion Loss [db] Return Loss [db] - Insertion Loss -5 Gb/s Spec -3 5 Gb/s Spec -35 3m 3AWG Return Loss -3 Gb/s Spec 5 Gb/s Spec -35-3m 3AWG Insertion Loss Deviation [db] ICN [mv, RMS] Insertion Loss Deviation Gb/s Spec 5 Gb/s Spec 3m 3AWG INTEGRATED CROSS-TALK NOISE IEEE / OIF Limits 3m 3AWG IL at.9 GHz [db] Figure : Gb/s IEEE P8.3ba spec and its extrapolation to 5 Gb/s. Shown are Insertion Loss/attenuation specification, Insertion Loss Deviation specification, Return Loss specification, and Integrated Crosstalk Noise specification.

13 Insertion loss [db] - -5 INSERTION LOSS - ATTENUATION -3 CEI 5G-LR Limits -35 3m 3AWG Return loss [db] 5 3 RETURN LOSS Insertion loss deviation [db] INSERTION LOSS DEVIATION CEI 5G-LR Limits 3m 3AWG CEI 5G-LR Limits 3m 3AWG Figure : Comparison to OIF CEI 5G-LR specification. Shown are fitted attenuation, Insertion Loss Deviation, and Return Loss. At this point, it is already obvious that only the cable passes the ICN requirements. The other cable assemblies do not meet the loss criteria for the ICN requirements. However, at the same time, they do meet the IL requirements of the spec. This points to a disconnect between the loss requirements in different parts of the OIF specification which will need to be reconciled... Addition of board-side connectors The next real component model to add to the overall cable assembly model was that for the board-side connectors. To obtain a connector Touchstone model appropriate for use at 5 Gb/s, a 3-D solid model of a QSFP Connector was analyzed using CST MicroWave Studio. Figure 3 shows the solid model. Note that it includes the connector footprint consisting of surface-mount (SMT) pads on the PCB side and card-edge pads (not shown) on the paddle card side. After cascading the connector model onto both ends of the cable assembly model, all of the performance parameters were plotted against the extrapolated IEEE P8.3ba limits in Figure and against the OIF limits in Figure 5. 3

14 Figure 3: 3-D model of QSFP board-side connector. Insertion Loss 8 Insertion Loss Deviation Insertion Loss [db] Return Loss [db] - -3 Gb/s Spec 5 Gb/s Spec - 3m 3AWG Return Loss - Gb/s Spec 5 Gb/s Spec 3m 3AWG Insertion Loss Deviation [db] ICN [mv, RMS] 6 - Gb/s Spec - 5 Gb/s Spec -6 3m 3AWG INTEGRATED CROSS-TALK NOISE IEEE / OIF Limits 3m 3AWG IL at.9 GHz [db] Figure : Gb/s IEEE P8.3ba spec and its extrapolation to 5 Gb/s. Shown are Insertion Loss/attenuation specification, Insertion Loss Deviation specification, Return Loss specification, and Integrated Crosstalk Noise specification

15 Insertion loss [db] - -5 INSERTION LOSS - ATTENUATION -3 CEI 5G-LR Limits -35 3m 3AWG Return loss [db] 5 3 RETURN LOSS Insertion loss deviation [db] INSERTION LOSS DEVIATION CEI 5G-LR Limits 3m 3AWG CEI 5G-LR Limits 3m 3AWG Figure 5: Comparison to OIF CEI 5G-LR specification. Shown are fitted attenuation, Insertion Loss Deviation, and Return Loss. As can be seen from Figure, only the cable passes the IL requirement for the IEEE spec. However, the impedance mismatches introduced by the connector now causes all of the cable assemblies to fail the IEEE ILD spec beyond GHz, and the IEEE RL spec beyond 7 GHz. The different shape of the IL curve fit also permits only the 3m 6 AWG cable assembly model to pass the OIF IL requirement, as can be seen from Figure 5. Again, the impedance mismatches causes all cable assemblies to fail ILD spec beyond GHz and RL spec beyond 7 GHz...5 Addition of test-board-fixture traces The final real component model to add to the overall cable assembly model was that for the test-board fixture traces. These models were generated by using a quasi-static -D analysis tool that is proprietary to FCI used earlier in section... The length of the traces was chose to meet the specification defined in section of the IEEE P8.3ba Draft 3. document. This amounted to approximately 3 cm of traces for each fixture. The board material used for the fixtures was FR, with a dielectric constant of and a loss tangent of.. Microstrip traces were used on the fixtures. Cascading the Touchstone (S-parameter) model of the fixtures at each end of the model from section.. results in a cable assembly model that performs as shown in Figures 6 and 7 when compared against the extrapolated IEEE and OIF specs respectively. 5

16 Insertion Loss 8 Insertion Loss Deviation Insertion loss [db] Insertion Loss [db] Return Loss [db] Gb/s Spec 5 Gb/s Spec 3m 3AWG Return Loss - Gb/s Spec 5 Gb/s Spec 3m 3AWG ICN [mv, RMS] Insertion Loss Deviation [db] 6 - Gb/s Spec - 5 Gb/s Spec -6 3m 3AWG INTEGRATED CROSS-TALK NOISE CEI 5G-LR Limits 3m 3AWG IL at.9 GHz [db] Figure 6: Gb/s IEEE P8.3ba spec and its extrapolation to 5 Gb/s. Shown are Insertion Loss/attenuation specification, Insertion Loss Deviation specification, Return Loss specification, and Integrated Crosstalk Noise specification INSERTION LOSS - ATTENUATION -3 CEI 5G-LR Limits -35 3m 3AWG Insertion loss deviation [db] INSERTION LOSS DEVIATION CEI 5G-LR Limits 3m 3AWG

17 Return loss [db] 5 3 RETURN LOSS CEI 5G-LR Limits 3m 3AWG Figure 7: Comparison to OIF CEI 5G-LR specification. Shown are fitted attenuation, Insertion Loss Deviation, and Return Loss. As can be seen from Figures 6 and 7, only the cable passes the IL requirement for the IEEE spec while none of the cables pass the OIF IL spec. The impedance mismatches introduced by the various parts of the cable assembly now cause all of the cable assemblies to fail the IEEE ILD spec beyond GHz and the OIF ILD spec beyond GHz, but are not severe enough to cause a RL failure when compared to either spec. None of the cables pass the ICN requirements. 5. Eye patterns, BER simulations, and the addition of signal conditioning Having characterized the passive channel composed of the cable assembly model and having observed its performance compared against the two specifications, it is now useful to perform eye pattern and BER simulations to see how the channel actually performs when transmitting a signal with a data rate of 5 Gb/s. It also makes sense to apply commonly-used signal-conditioning measures to the channel to see how the performance could be further enhanced. Transceiver parasitics, as described in section 3, have been included in the eye pattern and BER analysis. All simulations used the standard QSFP transmit- and receive-pair configuration shown in Figure 8. Rx Rx Tx Tx Rx Rx Tx Tx Figure 8: TX/RX configuration as defined by the QSFP specification. Figures 9 to 7 show the eye patterns, bathtub curves and BER eyes for the various cable assembly models at 5 Gb/s. The simulations include the multiple reflections caused by the transceiver parasitics. The figures also show the results of implementing signal conditioning. In order to re-open the eyes, the transmitter uses a PE filter with npr pre-cursor taps and npo post-cursor taps. The receiver uses a DFE with n baud-spaced taps. Three different settings for the transmitter and receiver taps have been studied:. 3-taps TX PE ( pre-cursor, post cursor) + 5 baud-spaced RX DFE taps. 3-taps TX PE ( pre-cursor, post cursor) + 7 baud-spaced RX DFE taps 3. 5-taps TX PE ( pre-cursors, post cursors) + 7 baud-spaced RX DFE taps 7

18 At each length the PE and DFE taps are set so that the relevant number of pulse response pre-cursors and post-cursors are made zero. In case of an n-taps DFE, the tap values are set in such a way that the n first post-cursors are zero. In case of a PE with npr pre-cursor taps and npo post-cursor taps, the tap values are set in such a way that the first npr precursor taps and the first npo post-cursor taps are zero. The sum of the magnitudes of all PE tap values is equal to HEIGHT = 65 mv -.8 WIDTH =.36 ns JITTER =. UI HEIGHT = 95 mv -. WIDTH =. ns JITTER =. UI BER Sampling point [UI] Figure 9: Input eye, output eye, bathtub curve, and BER eye for the 3m 3AWG ParaLink -3 cable assembly at 5 Gb/s with 3-tap TX PE and 5 baud-spaced RX DFE taps HEIGHT = 65 mv -.8 WIDTH =.36 ns JITTER =. UI HEIGHT = 97 mv -. WIDTH =.3 ns JITTER =. UI

19 BER Sampling point [UI] Figure : Input eye, output eye, bathtub curve, and BER eye for the 3m 3AWG ParaLink -3 cable assembly at 5 Gb/s with 3-tap TX PE and 7 baud-spaced RX DFE taps HEIGHT = 53 mv -.8 WIDTH =.35 ns JITTER =. UI HEIGHT = 3 mv -. WIDTH =.7 ns JITTER =.33 UI BER Sampling point [UI] Figure : Input eye, output eye, bathtub curve, and BER eye for the 3m 3AWG ParaLink -3 cable assembly at 5 Gb/s with 5-tap TX PE and 7 baud-spaced RX DFE taps. 9

20 HEIGHT = 6 mv -.8 WIDTH =.35 ns JITTER =. UI HEIGHT = 75 mv -. WIDTH =. ns JITTER =.8 UI BER Sampling point [UI] Figure : Input eye, output eye, bathtub curve, and BER eye for the ParaLink -3 cable assembly at 5 Gb/s with 3-tap TX PE and 5 baud-spaced RX DFE taps HEIGHT = 6 mv -.8 WIDTH =.35 ns JITTER =. UI HEIGHT = 79 mv -. WIDTH =. ns JITTER =.5 UI

21 BER Sampling point [UI] Figure 3: Input eye, output eye, bathtub curve, and BER eye for the ParaLink -3 cable assembly at 5 Gb/s with 3-tap TX PE and 7 baud-spaced RX DFE taps HEIGHT = 6 mv -.8 WIDTH =.3 ns JITTER =. UI HEIGHT = 85 mv -. WIDTH =.6 ns JITTER =.35 UI BER Sampling point [UI] Figure : Input eye, output eye, bathtub curve, and BER eye for the ParaLink -3 cable assembly at 5 Gb/s with 5-tap TX PE and 7 baud-spaced RX DFE taps.

22 HEIGHT = 565 mv -.8 WIDTH =.37 ns JITTER =.8 UI HEIGHT = 76 mv -. WIDTH =.6 ns JITTER =.35 UI BER Sampling point [UI] Figure 5: Input eye, output eye, bathtub curve, and BER eye for the ParaLink -3 cable assembly at 5 Gb/s with 3-tap TX PE and 5 baud-spaced RX DFE taps HEIGHT = 565 mv -.8 WIDTH =.37 ns JITTER =.8 UI HEIGHT = 8 mv -. WIDTH =.6 ns JITTER =.3 UI

23 BER Sampling point [UI] Figure 6: Input eye, output eye, bathtub curve, and BER eye for the ParaLink -3 cable assembly at 5 Gb/s with 3-tap TX PE and 7 baud-spaced RX DFE taps HEIGHT = 55 mv -.8 WIDTH =.37 ns JITTER =.8 UI HEIGHT = 77 mv -. WIDTH =.8 ns JITTER =.9 UI BER Sampling point [UI] Figure 7: Input eye, output eye and bathtub curve for the ParaLink -3 cable assembly at 5 Gb/s with 5-tap TX PE and 7 baud-spaced RX DFE taps. Figure 8 shows a summary table of the various eye openings achieved when using the compliant transceiver described in section 3, to transmit a 5 Gb/s signal through the complete cable assembly model, described in section..5. Note that transceiver 3

24 parasitics were included in the simulations. As can be clearly seen, all signal conditioning implementations re-opened the eyes. Signal Conditioning Used 3-tap TX PE, 5-tap RX DFE 3-tap TX PE, 7-tap RX DFE 5-tap TX PE, 7-tap RX DFE 3m 3AWG Cable Cable Cable Eye opening Eye Height, Eye opening Eye Height, at BER of Eye Width at BER of Eye Width e (mv, ps) e (mv, ps) Eye Height, Eye Width (mv, ps) 95 mv, ps 97mV, 3 ps 3mV, 7ps.8 UI.35 UI. UI 75mV, ps 79 mv, ps 85 mv, 6 ps. UI.8 UI.5 UI 76 mv, 6ps 8 mv, 6ps 77 mv, 8ps Eye opening at BER of e Figure 8: Eye openings of various cables when transmitting data at 5 Gb/s using different signal conditioning schemes..55 UI.58 UI.65 UI A number of Gb/s receivers in the market require a minimum eye opening of. UI at a BER of e in order to successfully receive the transmitted data. As can be seen from Figure 8, the signal conditioning scheme using a 5-tap TX PE and 7-tap RX DFE was able to successfully open the eyes to more than. UI at a BER of e in all the cables assemblies studied, when transmitting at 5 Gb/s. This suggests that even with existing receiver sensitivities, one could hope to successfully receive data at 5 Gb/s through a copper cable assembly. If receiver capabilities increase in the future, even a less complex signal conditioning scheme could possibly be used. 6. Conclusions In the effort to evaluate the transmission of signals at a data rate of 5 Gb/s through passive copper cable assemblies, it was found that the feasibility of doing so depends upon the user s definition of success. If success is defined as meeting a specification, then fairly severe limitations are places upon the physical (length, gauge) and electrical (impedance match, loss, crosstalk) attributes of the cable assembly. Alternatively, if success is defined as the achievement of actual open eye patterns and acceptable BER performance, then it is possible that the limitations on the attributes of the cable assembly are not as restrictive. Operating under the assumption that cable assemblies should meet the 5 Gb/s serial data transmission specifications as described in this paper, it is possible to determine a set of requirements for the constituent components comprising the overall assembly. 6. Limitations on raw cables The comparison of raw cables against the extrapolated IEEE and the OIF specifications placed limitations on the lengths and gauges of the cables which were different depending upon which spec was considered. According to the extrapolated IEEE spec, only the two shortest cables studied had sufficiently low loss to meet the IL requirement. This suggests that 3m is the maximum length that can be achieved for 5 Gb/s transmission using currently available cable technology.

25 The comparison against the OIF specification suggests however, that in addition to the two shortest cables, a cable of 5m length and 6 AWG (which almost passed the IEEE spec) also met the requirements. Regardless of which specification is being compared against, it was found that only the curve fit defined in the IEEE P8.3ba specification is appropriate for the evaluation of cable assemblies owing to its more accurate representation of the loss characteristics of a cable. 6. Limitations after adding paddle cards The addition of lossy paddle card traces to each end of the raw cables did not change the results stated above for the raw cables in the case of either specification. However, it should be noted that the impedance of the traces was well matched (i.e. essentially perfect,) and hence did not have a detrimental effect on any of the characteristics observed. It might be useful in the future to study the effects on performance of less than well matched paddle card traces. 6.3 Limitations after adding wire terminations The addition of the wire-to-paddle card terminations introduced crosstalk for the first time into the analysis. This did not cause any additional cables to fail IL, ILD or RL requirements, but it did narrow down to a single cable assembly, those that could pass the ICN requirement. Specifically, only the 3m 6 AWG cable assembly passed all of the extrapolated IEEE requirements. When compared against the OIF requirements, the results and conclusion were exactly the same in that only the 3m 6 AWG cable assembly passed. 6. Limitations after adding board-side connectors The addition of board-side connector models to the assembly model changed the concavity of the Insertion Loss curve fits. This can be attributed to the loss characteristic of the connector which tends to have a shape resembling that of a low-pass filter and apparently was able to dominate the overall shape of the IL curve.. Despite this, the same cable assembly as before passed the IL requirement for the IEEE spec. However, the impedance mismatches introduced by the connector now caused all of the cable assemblies to fail the ILD spec beyond GHz, and the RL spec beyond 7 GHz. The different shape of the IL curve fit also permitted only the 3m 6 AWG cable assembly model to pass the OIF IL requirement. Again, the impedance mismatches caused all to fail ILD spec beyond GHz and RL spec beyond 7 GHz. 6.5 Limitations after adding test-board-fixture traces The addition of the test-board-fixture traces caused all the cable assemblies to fail the OIF IL spec. Only the cable passed the IL requirement for the IEEE spec The impedance mismatches introduced by the various parts of the cable assembly now caused all of the cable assemblies to fail the IEEE ILD spec beyond GHz and the OIF 5

26 ILD spec beyond GHz. but were not severe enough to cause a RL failure when compared to either spec. Concerning ICN, which applies identically to both specs, none of the cable assemblies met the requirements. Interestingly, the reason for failure of some cable assemblies under consideration was not due to excessive crosstalk, but rather, was due to excessive loss. It might be possible to modify the ICN limits in a reasonable manner to account for the additional loss to be expected in cable assemblies. One possible extension to the limit which would permit more cable assemblies to pass is shown in Figure 5. This assumes, of course, that a transceiver would exist that would be able to tolerate or compensate for the additional loss. ICN [mv, RMS] 8 6 INTEGRATED CROSS-TALK NOISE CEI 5G-LR Limits 3m 3AWG IL at.9 GHz [db] Figure 5: Possible extrapolation of the ICN limit to permit cable assemblies with higher loss to pass. 6.6 Eye patterns and bathtub curves Knowing that technology continues relentlessly to improve, it is probably not reasonable to assert that 5 Gb/s transmission over copper cable assemblies is not feasible just by virtue of the fact that the assemblies considered here did not meet either of the proposed set of specifications. Hence, it was demonstrated that reasonable eye openings and BER results could be achieved using currently-available signal conditioning technologies. Even the signal conditioning scheme with the least overhead (3-tap TX PE and 5-tap RX DFE) was able to open the eyes for all the cable assemblies considered. The fact that transmission was achieved despite failure to meet the specs clearly demonstrates the difficulty of creating specifications and requirements for passive channels at ever-increasing data rates. 7. References [] IEEE Std 8.3ap -7, Annex 69B. [] IEEE P8.3ba Standard, Draft 3. [] C. Morgan, A Signal Integrity Comparison of 5 Gbps Backplane Systems Using Varying High-Density Connector Performance Levels, DesignCon 9. [3] CEI Implementation Agreement Draft X.X, Document OIF8.6.7, Oct 9. [] J. degeest, S. Sercu, Required Technologies for Transmission of x 5 Gb/s Over a Copper Backplane, DesignCon. [5] G. Oganessyan, M. Vrazel, Active Cable Interconnects for High-Speed Serial Communications, DesignCon. 6