抖动和眼图的视觉化分析
什么是抖动？
TIE 为信号相对于标准时钟或者标准信号的定时误差
TIE 在高速数字系统中即为抖动…抖动为实际数据与其理想位置的时间偏差0.0ns 0.990ns 2.000ns 2.980ns 4.000ns
-0.010ns 0.000ns -0.020ns TIE 0.000ns
什么是眼图？ 眼图是怎么形成的？
Random Jitter(随机抖动
•随机抖动符合高斯型分布•直方图(估计 ↔ pdf(数学模型 •抖动峰峰值=无穷大…无界！•内部热能现象•Flicker Noise, Shot Noise•热能的原子与分子振动•分子的解体•外部的宇宙射线
Deterministic Jitter(确定性抖动 •确定性抖动是非高斯分布并且有界
Periodic Jitter(周期性抖动 •TIE 随时间的变化是重复的、周期性的•Periodic jitter和相位调制(PM是等效的•系统时钟（抖动频率在MHz 量级）•开关电源（抖动频率在KHz 量级）
Duty Cycle distortion(占空比失真 •上升时间和下降时间不对称 •或者测试时参考电平选择不当0.0v
Inter-Symbol Interference(码间干扰抖动
•DDJ 或PDJ –数据相关性抖动或码型相关性抖动, 和ISI 的术语是等价的.
•码型是如何影响随后的比特位的？ ◦
由于传输链路的效应、反射等
换个角度看抖动，时域
看看我们有了什么视角？
7/7/2016
11抖动视觉化–时间趋势图
▪直方图告诉了我们分布，但是只有统计特性，缺少了时间信息▪时间趋势图可以直观告诉我们波形里是否有特定频率的调制▪下图为5个周期
SSC @ 30khz
抖动视觉化
Gaussian Random Noise Sinusoidal Jitter 7/7/201612
7/7/2016
13抖动视觉化–频谱图 ▪从频域上观测抖动 ▪
抖动中决定性的频率成分会在谱线上明显超出噪底
哪个眼图好？哪个直方图好？ 视觉化眼图和抖动的问题？
浴盆曲线
误码率是关键vs. UI张开程度
•For a given position in the time there’s a given probability of error –“BER ”, Bit Error Ratio
•1 UI
基于示波器分析的浴盆曲线 Rj δδ/Djδδ与Tj @ BER
Assume bi-modal distribution (dual-Dirac, measure Tj at two BER
抖动类型分析 •
抖动分离为误码产生的根本原因提供了更精确的定位和分析方法•抖动分析方法，参照T11 MJSQ，已经被工业界广泛接受Constituent Components of Jitter
= Unbounded = Bounded
7/7/201618Jitter Visualization –Bathtub Plot
▪▪Note the eye closure of System I vs. System II due to the RJ-RJ is unbounded so the closure increases as BER level increases▪System I has .053UI of RJ with no PJ
▪System II has .018UI of RJ and .14UI of PJ @ 5 and 10Mhz System ISystem I System II System II
Tektronix -Innovators of Jitter Analysis
•1998First Real-Time Scope Based Jitter Analysis Software
•2002 Invented SW Based PLL Clock Recovery and the Spectral Approach for Jitter Separation•2004–Invented RT Eye rendering on a Real Time Scope
•2004-First vendor to support both modeled (Dual-Dirac and measured (Spectral jitter methods•2005-Invented measurements with Jitter and Noise reconciliation
•2011-First scope vendor with BUJ support •2015–
RT Noise Analysis and Sampling BER and PDF Mask Testing
抖动和眼图的视觉化 眼图怎么切割的？时钟决定！
TIE 抖动需要参考时钟 •
参考时钟提取的过程就是时钟恢复•参考时钟有几种确定的方式: ◦Constant Clock with Minimum Mean Squared Error This is the mathematically “ideal” clock
But, only applicable when post-processing a finite-length waveform Best for showing very-low-frequency effects
Also shows very-low-frequency effects of scope’s timebase ◦Phase Locked Loop (e.g. Golden PLL
Tracks low-frequency jitter (e.g. clock drift
Models “real world” clock recovery circuits very well ◦Explicit Clock
The clock is not recovered, but is directly probed ◦Explicit Clock (Subrate
The clock is directly probed, but must be multiplied up by some integral factor 7/7/2016 21
Importance of Clock Recovery
•From spec, “The jitter measurement device shall comply with the JTF”. •How do I verify JTF?
◦JTF is difference between input clock (ref and input clock (unfiltered
◦Use 1100b or 0011b pattern (proper 50% transition density ◦Check 1 LF attenuation, 2 -3 dB corner frequency, and 3 slope 7/7/201622
JTF vs PLL Loop Bandwidth
•Configuring the correct PLL settings is key to correct measurements
•Most standards have a reference/defined CR setup ◦For example, USB 3.0 uses a Type II with JTF of 4.9Mhz •Type I PLL
◦Type I PLL has 20dB of roll off per decade ◦JTF and PLL Loop Bandwidth are Equal •Type 2 PLL
◦Type II PLL has 40dB of roll off per decade ◦JTF and PLL Loop Bandwidth are not Equal
▪For example, USB 3.0 uses a Type 2 PLL with a JTF of 4.9Mhz. The corresponding loop bandwidth is 10.126 Mhz
▪Setting the Loop Bandwidth as opposed to JTF will lead to incorrect jitter measurement results
7/7/201623
PLL Loop Bandwidth vs. Jitter Transfer Function
(JTF
A: Constant Clock Recovery B: PLL Clock Recovery Ratio of B/A 7/7/2016 24
JTF Filtering Effects based on different PLL bandwidths
7/7/2016
27f 3dB = 30 kHzf 3dB = 300 kHzf 3dB = 3 MHz
Jitter for Busy PeopleHints, Tips and Common Errors
Using the Jitter Analysis Tools
•Issues manifested in different layers of the protocol stack
◦Crosstalk, jitter, reflections, skew
◦Disparity, encoding or CRC errors •Where do I start debugging? •Jitter and Eye Diagram Tools ◦Oscilloscope-based for quick results ▪Fast jitter measurements with ▫‘One Button’ Jitter Wizard
▪Compare timing, jitter, eye, amplitude measurements
▪User-definable clock recovery, filters, pass/fail limits, and reference levels
More Hints for Successful Jitter Analysis
•Clock Recovery has a great deal of influence on jitter results. Think about what you’re trying to accomplish.
◦Constant-Clock is the most “unbiased”
Often best if you’re trying to see very-low-frequency effects But it can also show wander in the scope’s timebase ◦PLL recovery can model what a real data receiver will see
It can track and remove low-frequency effects, allowing you to “see through” to the jitter that really contributes to eye closure
◦Explicit-Clock is appropriate if your design uses a forwarded clock
Make sure your probes are deskewed
Hints for looking at Spread-Spectrum Clock
•If you want to see the SSC effects, use TIE and PLL clock recovery with a bandwidth of at least 1 MHz. A Type-II (2nd -order PLL will track out the SSC more effectively than a Type-I PLL.•want to observe the SSC profile:
◦Use a Period measurement and turn on a 3rd -order low-pass filter(in DPOJET with a bandwidth of 200 kHz
Because Period trends accentuate high frequency noise, the low-frequency SSC trend will be obscured if you don’t use a filter You can’t use a Frequency measurement directly. The combination of filtering and the reciprocal operation (Freq = 1/Per cause distortion in the resulting waveshape. (This is a mathematical fact, not a DPOJET defect.
◦If you use a TIE measurement, you’ll see modulation that looks like a sine wave. This is normal. It’s because TIE measures phase modulation, which is the integral of frequency. It turns out that the integral of a triangle wave looks very much like a sine wave.
误码率与噪声分析
Anatomy of a Serial Data Link
Complete Link Channel
Aspirational goal: 0 errors
Practical Goal: Bit Error Rate < Target BER
•Since BER is the ultimate goal, why not measure it directly?
Serial Data Link Integrity = Bit Error Rate
•Bit Error Ratio Testers (BERTs are the tools for measuring BER directly•Why not use ONLY BERTs for Serial Data Link Analysis?
◦Difficult to model/emulate equalizer◦Measurements could take a very long time ◦Instruments are very expensive and not all that flexible ◦Does not analyze the root causes of the impairments of the links •Alternative approach: use a scope and advanced analysis tools ◦Easily move from Compliance to Debug
◦Better equipped to identify root causes of eye closure◦Equalizer can easily be modeled◦More cost effective◦Faster throughput
Why Measure Jitter and Noise?
▪Link Model: Transmitter + Channel + Receiver▪Transmitter generates a stream of symbols
▪Receiver uses a slicer to make a decision on the transmitted symbol▪The Bit Decision is made at a certain time (t of the symbol interval and a comparison of the sliced data to a threshold (v is performed▪Jitter impairs the time slicing position
▪Noise impairs the decision threshold ?
Jitter combined with Noise Analysis is a better predictor of BER performance!
A Quick Look at Jitter and Noise Duality
•Jitter analysis evaluates a waveform in the horizontal dimension based on when the waveform crosses ahorizontal reference line.
•Jitter decomposition is based on spectral analysis of Time Interval Error vs. time ◦Individual jitter components can be separated (i.e. PJ, RJ, DDJ, etc. ◦TJ can then be estimated at a target BER level
38
▪Noise evaluates along a vertical dimension on the basis of
crossings of a vertical reference line at some percentage of the unit interval (usually 50%.▪Noise decomposition is based on spectral analysis of voltage error vs. time
–Individual noise components can be separated (i.e. PN, RN, DDN, etc. –TN can then be estimated at a target BER level
抖动和噪声的解析 •
Jitter and Noise Decomposition provide deep insight into BER
Full Jitter Analysis vs. Mask Testing
•
statistical eye closure at any other voltage.
•Conventional mask testing considers both time and voltage , but cannot extrapolate eye closure at low BER.
Can we combine the best of both?
41Statistical Jitter + Noise Analysis
•By jointly analyzing Jitter and Noise, behavior at allpoints in the eye can be extrapolated at low BER
•The methodology is analogous to current jitter analysis, but is performed across both dimensions of the eye
◦Jitter and noise are separated into components (Random, Periodic, Data-Dependent,…◦The components are reassembled into a model that allows accurate extrapolation.
42Timing-Induced Jitter
•Since jitter is defined as a shift in an edge’s time relative to its expected position, it is
easy to think of jitter as being causedby horizontal (chronological
displacement.•Note that the displaced edge (green has not moved vertically in this example.
43Noise-Induced Jitter
•Consider a burst of voltage noise (right that displaces a waveform vertically.◦In this case, the displaced edge (green has not moved horizontally.
•The jitter as measured at the chosen reference voltage is identical in these cases!◦So, why should we care?
•
Two fundamentally different effects have caused the same amount of jitter, and 44
Noise-to-Jitter (AM-to-PM Conversion •Since waveform transitions are never instantaneous, the slope (slew rate of the edge acts as a gain constant that controls how effectively noise is converted to
“observed jitter”.
45
Horizontal and Vertical Components of Random Jitter •We can think of RJ as being composed of two components. ◦Horizontally induced: RJ(h◦Vertically induced: RJ(v
•Since these two components are uncorrelated with each other, they add in the
RSS sense: RJ =RJ(h 2+RJ(v2
•Similarly, PJ can be decomposed into PJ(h and PJ(v based on root cause
46
Horizontal and Vertical Components of Random Noise
•We measure noise at a reference point in the bit interval (usually 50%•If slew rate isn’t zero, jitter (horizontal displacement causes observed noise
•So as with RJ, RN can be decomposed into components: ◦Horizontally induced: RN(h ◦Vertically induced: RN(v •
Similarly, PN can be decomposed into PN(h and PN(v based on root cause
Noise to Jitter and Jitter to Noise Conversion Consider: an “ideal” edge in a pattern actually has two impairments: ◦Jitter(h (see the blue trace–and Noise
(note that both of Jitter and Noise result in jitter on edgeThe Combined response (bottom
right includes the jitter caused bynoise Non-impaired bit edge We can separate the noise contribution of jitter for diagnostic purposes by breaking RJ into RJ(v and RJ(h
DPOJET and 80SJNB are the only tool that will show you this separation, and thus give you
an important troubleshooting hint: e.g. is it crosstalk causing trouble, or the clocks?
48Theory: Construction of the BER Eye •Consider a very simple pattern: 7 bit repeating
•Overlay multiple segments of the 7-bit pattern. Each one has noise and jitter, so although the bit pattern is clear, they follow many slightly different paths:
•Average many pattern repeats together. Everything that is uncorrelated with the pattern averages out. What remains is called the ‘correlated waveform’.◦This waveform fully characterizes DDJ, DCD, DDN, ISI –
all data dependent effects
49
Theory: Construction of the BER Eye –Part 2•The correlated waveform can be snipped into individual bits and overlaid to form an eye diagram, using the recovered clock as the alignment reference. This forms the ‘correlated eye
’:
50
Theory: Construction of the BER Eye –Part 3
•Spectral jitter separation is used to find PDFs of the random and periodic jitter.•The RJ and PJ PDFs are convolved to find the uncorrelated jitter PDF (red •A similar analysis of the noise yields the uncorrelated noise PDF (blue ◦Care must be taken to properly account for AM-to-PM and PM-to-AM conversion in these steps; otherwise some noise or jitter would be ‘double-counted’.
•Two-dimensional convolution is used to create a joint PDF of uncorrelated jitter + noise. (We can call this the ‘jitter/noise set’
51Theory: Construction of the BER Eye –Part 4
•The jitter/noise set is convolved (two-dimensionally with the correlated eye for the ‘1’ bits to get the overall
(correlated + uncorrelated PDF for ‘1’ bits
52Theory: Construction of the BER Eye –Part 5
•The ‘1’ bit PDF is integrated vertically (from bottom to top to get the ‘1’ bit CDF (Cumulative Distribution Function
◦In this color-graded view, each color represents a particular BER level
53Theory: Construction of the BER Eye –Part 6
•
A similar treatment for ‘0’ bits yields the ‘0’ bit CDF
54Theory: Construction of the BER Eye –Conclusion
•The ‘1’ bit and ‘0’ bit CDFs are added to get the overall “BER Eye”
◦A particular BER contour can be found in the 3D version of this plot by slicing it horizontally, or by extracting a specific color on either version
◦Since this ‘eye’ looks rather unconventional, DPOJET extracts the BER contours and then overlays them with the rendered eye.3D View Color-Graded View
因篇幅问题不能全部显示，请点此查看更多更全内容