Challenges and Solutions in Characterizing a 10 Gb Device

Transcription

1 Challenges and Solutions in Characterizing a 10 Gb Device DesignCon January 28 th, 2013 Brian Fetz Daniel Rubusch Rob Sleigh 1

2 Topics Industry Demands Difficulties in High Speed Digital Design Measurement Instruments Conflict? Topologies of Jitter Measurement Devices Considerations in Jitter Measurement Correlation Demos and Discussion Crosstalk considerations Summary

3 Industry Demands Smart phone connectivity Tablet Revolution Do you think this has any effect on the network infrastructure? Expectations: Leveling out or Continued Increase of Rates? Compression techniques? PAM 4? In the meantime. We sweat!

4 High Speed Interfaces (Rates by Year by Industry) Infiniband CEI 56G Thunderbolt (25G) GbE, CEI 28G VSR PCIe4 (16Gb) 10 Thunderbolt (10G) SAS3 (12 Gb) 40GbE SATA4 (12 Gb) USB 3.5 (10Gb) ) MIPI M-PHY (10/11.6 Gb) Speed (Gb) SAS2 (6 Gb) SATA3 ( 6 Gb) PCIe3 (8Gb) FB-DIMM (4.8 Gb-9.6 Gbx24) HDMI 2.0 ( 6 Gb/s) edp 1.2 (5.4 Gb) DP 1.3 (8.1 Gb) GDDR5 (7Gbx32) UFS 2.0 (6Gbx2) edp 1.5? (8.1 Gb) 5 DP 1.2 (5.4 Gb) M-PHY (5/ 5.8 Gb) USB 3 (5Gb) PCIe2 (5Gb) MIPI (CSI-3, LLI) M-PHY (5/5.8 Gb) 3 edp 1.1 (2.7 Gb) HDMI 1.4 (3.4 Gb/s) MHL 3 Gb DDR4 ( Gbx8) HBM (2Gbx128) UFS 1.1 (3Gb) 2 1 DDR3 (800MB-2.3 Gbx8) MIPI (DigRFv4) / M-PHY (1.25/1.46 Gb) W-USB 1.1(1Gb) MIPI (CSI-2, DSI 1.1) D-PHY (1G/ 1.5G) USB 2.0(480 Mb) MIPI (DigRFv4, UFS) / M-PHY (2.5/ 2.9 Gb) Wide I/O2 (533Mbx512) Year

5 Case for Measurements: (consistent and accurate) 1 Gbs=> Unit Interval = 1000 ps 10 Gbs=> Unit Interval= 100 ps 28 Gbs=> Unit Interval= 36 ps ps Anyone feeling squeezed?

6 But its just not even THAT simple + Tx - Emp Txp Txn Connector Channel Connector Rxp Rxn EQ + Rx - TP1 TP2 TP3 TP4 Closed Eye Eye Opened with Equalization

7 Above 4 GHz Everything is out to get you!! Material Layout Transmission Lines Parasitics Interconnects Crosstalk Humidity Power Planes Temperature Components Tx Noise Pre-emphasis Delay Ground Bounce ISI Skew Frequency Response Crosstalk Reflections Skew Noise Match Equalization modeling Clock Recovery/PLL Performance Skew/Common Mode Sensitivity

8 This image cannot currently be displayed. This image cannot currently be displayed. Measurement Equipment The last thing you want is your measurement equipment to collude against you as well! You need confidence you have : --Right tools for the job --Right setup --Right connection --Right Expectation

9 The Objective: Support Specified Link BER Data Bit Error Ratio is the key metric Probability that a bit is received in error. Link Model (Tx, Channel, Rx) Transmitter Channel Receiver + Tx - Emp Txp Txn Cnctr Channel Cnctr Rxp Rxn EQ + Rx -

10 Evaluating Portions of the System Transmitter + Tx - Emp Txp Txn Cnctr Channel Cnctr Channel Cnctr Receiver Cnctr Rxp Rxn EQ + Rx -

11 Receiver Testing Jitter Components (RJ+DJ+ISI) N4903B JBERT RJ+DJ+ISI Frame Error Rec Rx Tx Loop Back Device Under Chip Test RJ DJ ISI

12 Source Testing + Tx - Emp Txp Txn Cnctr Voltage Level Eye Diagram Jitter and the components of jitter Skew Pre-Emphasis/De-Emphasis Frequency/Data Rate Spread Spectrum Clocking parameters

13 Jitter Components Total Jitter (TJ) Bounded UnBounded Deterministic Jitter (DJ) Random Jitter (RJ) Correlated with Data (DDJ) Uncorrelated with Data (BUJ) DutyCycle Distortion (DCD) InterSymbol Interference (ISI) Non Periodic (ABUJ) Periodic (PJ) Gaussians ( RJ RMS ) Tr, Tf Settling Time Xtalk Clocks Thermal Reflections Non flat Freq Response Non Linear CR Events Xtalk Shot 1/f Burst

14 Jitter measurements on the same DUT performed by two different instruments Sampling Scope Real Time Scope A seemingly simple jitter measurement on a 10 Gbps device Why can the results be so different?

15 Sampling vs Real-time Oscilloscopes Topologies, Pros and Cons Equivalent-time Sampling Oscilloscope Real Time Oscilloscope

16 Equivalent-Time Sampling Extremely wide bandwidths at low sample rates 2 4-1=15 bits 2 4-1=15 bits PRBS Pattern Trigger Trigger Point Sampling Point Sequential Delay Reconstructed Waveform A sample is taken, the data pattern repeats and the next sample is taken at a slight delay compared to the previous sample In practice, samples are very close together (can be less than 100 fs apart). Through multiple passes of the signal, the waveform can be precisely reconstructed

17 Sampling Scope Bandwidth is Independent of Sample Rate S Sampler input Measurement bandwidth is affected by how narrow the sampler control pulse is (can be just a few picoseconds) Sampler control pulse Since only one sample is taken, the A-D process can be very high resolution (up to 15 bits) with very low noise Sampler pulse: Low bandwidth High bandwidth

18 Eye Diagrams: Highly Synchronous Sampling at Arbitrary Bit Locations PRBS Re-Arm Time Trigger Point Sampling Point Clock Trigger Reconstructed Waveform One Bit MYTH!: Sampling scopes can only display repetitive signals

19 Real-time Sampling Sample entire waveform in one acquisition S(t) T T Trigger Event T=1/F S Could Trigger Here. Or Here Nyquist criterion obeyed: Fs > 2*BW of signal Interpolation is used to precisely fill in points in between actual sampled points to yield better resolution

20 Oscilloscope Block Diagram Comparison 8bits Real Time Oscilloscope 16 bits Sampling Oscilloscope

21 Platform Pros and Cons Sampling Oscilloscope Bandwidth to > 90 GHz Lowest noise & jitter High precision long term view Precision optical receivers Price Real Time Oscilloscope Bandwidths to 63 GHz Sample rates up to 160 GSa/s High resolution short term capture Rich and flexible triggering Complete signal access (probing)

22 Jitter Measurements Same DUT Performed/Two Different Instruments Sampling Scope Real Time Scope A seemingly simple jitter measurement on a 10 Gbps device why can the results be so different?

23 What Causes Different Jitter Results? Clock Recovery Noise / Slew Rate Memory Depth Frequency Response Crosstalk Other issues to consider: 1. Jitter Thresholds (DCD) 2. Input Match (AC or DC coupled signal?) 23 S800 JITR_12_1 Labs.pptx January 28, 2013

24 Clock Recovery If you re not using equivalent clock recovery models, correlating jitter results will be VERY difficult. 24

25 Quick Review - Clock Recovery (CR) Basics o o o Provides a recovered clock for receiver Manages jitter in the system Standards specify CR Phase Locked Loop (PLL) order, bandwidth, peaking, or damping factor Input Signal Data Input Phase Detector Phase Error Amplifier Voltage Controlled Oscillator (VCO) Recovered Clock Sampler (Receiver) Data relative to a clean clock (narrow loop BW) OR Data relative to recovered clock (wide loop BW) Basic CR Block Diagram Narrow CR Loop BW OR Wide CR Loop BW PLL Jitter Transfer Function (JTF) indicates how much of the jitter on the input signal is transferred to the recovered clock (output) low-pass filter function (LPF) JTF Closed loop gain out A( s) G( s) G( s) e in 1 A( s) j ( s) Jitter Multiplier Clock Recovery PLL Response Jitter Transfer Function (JTF) and Observed Jitter Transfer Function (OJTF) 0 1.0E E E+3 1.0E E E+6 Frequency (Hz) Observed Jitter Transfer Function (OJTF) indicates the jitter that is observed by the receiver (scope) high frequency jitter on the data stream is transferred to the receiver (HPF) OJTF 1 JTF 1-G( s) 1 G( s) e j ( s) BEWARE of Clock Recovery (PLL) Definitions! Standards (and scopes) describe PLL requirements differently. Agilent 86100C/D Sampling Scope CR loop BW setting configures JTF JTF Example: Ethernet, SONET/SDH Agilent 90K Series Real-time Scope CR loop BW setting configures OJTF OJTF : SATA/SAS

26 Jitter Spectrum To understand how the CR PLL response impacts low frequency jitter, it is useful to observe jitter in the frequency domain Magnitude Frequency Offset frequency

27 Jitter Spectrum Shows distribution of low frequency jitter and impact of clock recovery Narrow CR loop bandwidth Wide CR loop bandwidth Spectral lines indicate deterministic jitter (including SSC and its odd harmonics) Jitter floor (without tones) is random jitter Observe all incoming jitter Track out low frequency jitter Clock Recovery response greatly impacts amount of jitter seen by receiver, and/or measured by an oscilloscope!

28 Clock Recovery Models 1 st Order PLL: JTF BW = OJTF BW Peaking/DF = none Roll-Off: 20 db/decade - Less ability to track out low frequency jitter and stay locked - Real hardware CR does not behave this way HW CR Loop Response Jitter Spectrum 2 nd Order, Type 2 PLL: Bandwidth: JTF BW > OJTF BW Peaking/Damping Factor: need to specify Roll-Off: 40 db/decade (tracks out low jitter more than 1 st order PLL) HW CR response may have higher peaking in OJTF than desired. This will amplify jitter in this region. Note significance depends on DUT jitter spectrum. 3 rd Order PLL: JTF BW > OJTF BW - Specify zero, gain, pole frequencies. - Roll-Off: 60 db/decade below zero frequency - Use PLL Response Tutorial workbook to model. Desired SW CR Loop Response e.g. match a standard exactly Jitter Spectrum Less Peaking 86108B FTD_DCA_224 Agilent Restricted March 2012

29 Jitter Spectrum Analysis and SW Clock Recovery Emulation using Agilent 86100D/86108B-JSA 86108A/B Module Device Under Test Data or Clock Signal ( Jitter Filter ) Integrated Hardware Clock Recovery Filtered Signal ( Jitter Fitler ) Ideal Software Clock Recovery Emulation Real CR PLL response Adjustable Loop Bandwidth Adjustable Peaking (discrete) Ideal, flexible CR PLL response Adjustable Loop Bandwidth Adjustable Peaking (continuous) Desired SW CR Loop Response e.g. match a standard exactly Less Peaking HW CR response may have higher peaking in OJTF than desired. Jitter amplification will occur in region where unwanted peaking exists. Jitter Spectrum Apply ideal PLL using Software Clock Recovery Emulation Jitter Spectrum Note how much of an increase depends on DUT jitter spectrum. Higher Accuracy Hardware only clock recovery Ideal SW Clock Recovery Model

30 Clock Recovery Comparison Always use similar clock recovery models Apples-to-Apples setup Agilent K X-Series X-Series Agilent 86100D with 86108B 30

31 10 Gb/s Jitter Measurement Demo Part 1 Perform a jitter measurement using 10 MHz CR loop bandwidth. 10G Pattern Generator D+ D- Agilent 86100D CR BW: 10 MHz (JTF) Agilent 90K X-Series CR BW: 10 MHz (OJTF) Not an apples-apples measurement with respect to clock recovery!

32 10 Gb/s Jitter Measurement Demo Part 2 Perform a jitter measurement using 2 nd Order CR response with 10MHz OJTF and DF. 10G Pattern Generator D+ D- JTF: 2 nd Order, 20 MHz Loop BW, 2dB Peaking OJTF: 2 nd Order 10 MHz Loop BW, DF OJTF: 2 nd Order 10 MHz Loop BW, DF We are using the same CR setup now, but are there other things we should look at?

33 Clock Recovery: Clock-to-Data Trigger Delay Clock-to-Data Trigger delay adds phase shifts Effect on OJTF => peaking (amplifies jitter!) Introduced when using External Clock Recovery JTF (BERT, Sampling Scopes, Explicit Clocks ) PLL+DCA out ( s) ( s) Peak is reduced from ~1.4 down to ~ 1.1 by decreasing the effective trigger delay in. e -j (trigger delay) 24 ns delay 0 ns delay Delay data signal using specially matched cables (see vendo (degrades signal integrity, however)

34 Oscilloscope Noise Impacts Measured Jitter - Demo Measure AC rms measurement at proper Volts/Div scale for DUT signal Agilent 86100D/86108B Series: ~ 640 uv (at 35 GHz BW Setting) Agilent 90K X-Series: ~ 6.1 mv (at 137 mv/div and 32 GHz BW Setting) Note - single-ended noise measurements since we re performing a comparison using singleended signals (analyzing P and N from the same DUT)

35 Manually Determine Induced Jitter due to Scope Noise and Signal s Slew Rate RN = Random Noise(rms) Slew Rate = rate of change of signal in V / ns = Delta V/ Delta T Induced Jitter due to scope noise: D / 86108B DCA-X Noise = 640 uv Slew Rate = mv / 8.34 ps = 20.8 V/ns Induced Jitter = RN / SlewRate = 640uV / 20.8V/ns Induced Jitter = 31 fs Delta T Delta V 2. 90K X-Series Oscilloscope Noise = 6.1 mv Slew Rate = 26 V/ns Induced Jitter = RN /SlewRate = 6.1mV / 26 V/ns Induced Jitter = 234 fs The faster the edge, the smaller the problem! And vice-versa!

36 Estimate Jitter due to Intrinsic Scope Jitter/Noise and Signal s Slew Rate (AM-to-PM Conversion) Example: 86100D / 86108B 1. DUT Random Jitter = 200 fs 2. Scope Random Jitter = 50 fs Random Timing Jitter = 206 fs = SQRT [(200^2)+(50^2)] 3. Noise Induced Jitter from scope = 31 fs (see previous page) Example: 90K X-Series 1. DUT Random Jitter = 200 fs 2. Scope Random Jitter = 75 fs Random Timing Jitter = 213 fs = SQRT [(200^2)+(75^2)] 3. Noise Induced Jitter from scope = 234 fs (see previous page) Measured Jitter = SQRT [(Timing Jitter)^2 + (AM-to-PM Jitter)^2)] Measured Jitter = SQRT [(206)^2 + (31)^2)] = 208 fs Measured Jitter = SQRT [(213)^2 + (234)^2)] = 317 fs Scope jitter results include noise induced jitter (AM-to-PM conversion). Results change due to signal slew rate and random noise.

37 Use FlexDCA simulator to determine induced jitter due to noise/slew rate on saved waveforms Actual saved waveform => Playback and add 640 uv rms random noise Add Noise Only Add Jitter + Noise DCA 35GHz BW, 50 mv/div RJ due to scope noise only. RJ due to timing jitter and scope noise. Minimize scope BW to reduce noise, especially if signal has slow edges!

38 Summary - Noise / Slew Rate As random noise (RN) increases, random jitter increases. Especially problematic with slower edge speeds! Minimize oscilloscope noise. Use only enough BW to capture signal. 38 S800 JITR_12_1 Labs.pptx January 28, 2013

39 Memory Depth and Wide vs Narrow Mode (RT Scope settings) Pay attention to the BER Bathtub plots! RJ vs Memory Depth: White Paper: (see page 35-37). Increase memory depth until RJ result stabilizes Check BER Bathtub curves to ensure a continuous function RJ Bandwidth: Narrow (pink) vs Wide (white) White Paper Start with narrow RJ setting Discontinuities in the BER Bathtub graph should be a red flag! 39 S800 JITR_12_1 Labs.pptx January 28, 2013

40 Receiver Frequency Response Differences in scope receiver s frequency response will change: Shape of eye diagram Amplitude Rise/fall times ISI 40 S800 JITR_12_1 Labs.pptx January 28, 2013

41 Frequency Response Step Response What does each filter response look like and do to a step? 33 GHz Sinc 50 GHz Sinc Filter Waveform: 10 GHz/div, 3 db/div Step Waveform: 100 ps/div Yellow frequency response Blue step response 50 GHz Flat Real-time Scope 50 GHz Bessel Real-time Scope Closest to Sampling Scope RX

42 Frequency Response 50 GHz BW 28 Gb/s eye with infinite TX bandwidth 86108B 50 GHz Raw Corrected to 50 GHz Sinc Which eye amplitude is higher? Answer: They re about the same. Corrected to 50 GHz Flat Corrected to 50 GHz Bessel

43 Frequency Response 33 GHz 28 Gb/s eye with infinite TX bandwidth 86108B 50 GHz Raw Corrected to 33 GHz Sinc Same input signal which eye amplitude is higher now? Answer: 33GHz Sinc, even though it has less BW (more energy above the mean 1 level.) Corrected to 33 GHz Flat Corrected to 33 GHz Bessel RX Frequency Response: Amplitude and Phase response are important.

44 Frequency Response Use FFT to determine how much BW to use 10Gbps signal measured using different BW on a RT Scope 32 GHz GHz 20 GHz 15GHz To ensure optimal waveform accuracy, use sufficient BW to capture all energy in the signal, but only use what you need to minimize noise.

45 Frequency Response 63 GHz Sinc Simulated 28G, 37 GHz 4 th Order Bessel 50 GHz Sinc 33 GHz Sinc Simulated Recommend using more than 33GHz BW for this 28Gb/s signal! 86108B 50 GHz Simulated 45 S800 JITR_12_1 Labs.pptx January 28, 2013

46 Agilent Application Software Take all the guess work out of setting up a scope for an accurate, and compliant, measurement D DCA-X N1012A CEI 3.0 6G/11G/25G/28G and 28G VSR (Draft) N1019A SFF-8431 (10G SFP+) N1019A User-Defined Application X-Series, Q-Series Scopes DDR, DVI, DisplayPort, Ethernet, Fully Buffered DIMM, GDDR, HDMI, MHL, MIPI, PCI Express, Secure Digital, Serial ATA I/II, Thunderbolt, USB, Wireless USB, XAUI, and more!

47 Agilent Application Software - Thunderbolt Demo TBT Application using TBT as DUT

48 DEMO Spectral RJ Extraction Tail Fit RJ Extraction

49 Crosstalk Bounded Jitter that is Deterministic and is uncorrelated with the Data may be Periodic or Aperiodic. Deterministic Jitter (DJ) Periodic is easily identified, Aperiodic jitter is not. Correlated with Data (DDJ) DutyCycle Distortion (DCD) InterSymbol Interference (ISI) Uncorrelated with Data (BUJ) Non Periodic (ABUJ) Periodic (PJ) The problem with crosstalk jitter arises from techniques to separate jitter into its constituent components Tr, Tf Settling Time Xtalk Clocks Reflections Non flat Freq Response Non Linear CR Events Xtalk Aggressor

50 Crosstalk cont The problem with crosstalk jitter arises from techniques to separate jitter into its constituent components ISI PJ By examining repetitions of the pattern you can definitively identify the ISI and remove it from the T.I.E record. RJ DJ ISI What remains is PJ and RJ and an easy way to separate them through spectral analysis---because periodic components are easily identified in the frequency domain RJ RJ time error likely to contain PJ PJ threshold 0 0 freq

51 Crosstalk (2) But what if crosstalk IS present? You can still remove the Data Dependent Jitter, but the PJ/RJ separation wont work because the spectrum of crosstalk MOST likely is NOISE LIKE. Therefore, Crosstalk (or Aperiodic Bounded Uncorrelated Jitter) will make the RJ look bigger time error likely to contain PJ RJ DJ ISI PJ threshold 0 0 freq You need to have a Gaussian Tailfit Extraction for the RJ

52 RJ Extraction with Crosstalk (ABUJ) Spectral vs. Gaussian RJ Extraction. No Crosstalk w/crosstalk X Spectral Extraction Examine slope continuity Spectral Extraction Compare actual Data with RJ estimates of both methods Gaussian Tailfit Extraction Gaussian Tailfit Extraction

53 Two Ways to Analyze ABUJ 1. Use Gaussian Tailfit Extraction 2. Two Pass Spectral Extraction Approach assumes you have control of the interferer assumes conveyed jitter of interferer is all BUJ 1.47 ps

54 ABUJ/Crosstalk Analysis 1. Gaussian Tailfit Extraction No interferer With interferer Victim Aggressor Aggressor at transition

55 ABUJ/Crosstalk Analysis 2. Two Pass approach a) Turn off crosstalk element(s). b) Measure jitter (jitter components) c) Turn on crosstalk element(s) d) Enter RJ rms value for RJ ( specify ) e) Crosstalk (ABUJ) will go into bounded portion of jitter which will prevent overestimation of RJ and Total Jitter ps

56 ABUJ/Crosstalk Analysis Two Pass Approach No interferer With interferer Victim Aggressor Aggressor at transition

57 Summary The engineer facing high speed problems is facing an uphill climb SI Problems Takes some effort to get correlation between real time scopes and sampling scopes. (Clock Recovery, frequency response, noise, etc.) Compliance packages can off load engineer from being an expert Don t be the mule! Offload it! Crosstalk may be present and if you suspect it, then it is helpful to have a couple of ways to understand its effects