|
ADC中的时钟抖动
本帖最后由 stupid 于 2011-4-8 09:52 编辑 2 Z- V! h% \' V* Q( a, g
; s$ Z% r; R, S7 z
Understanding the effect of clock jitter on high-speed ADCs (Part 1 of 2)
* k- t$ j8 e! L# eDerek Redmayne & Alison Steer, Linear Technology Corp.8/12/2008 12:44 PM EDTDigitizing high-speed signals to a high resolution requires careful selection of a clock that will not compromise the sampling performance of the analog-to-digital converter (ADC). In this article, we'll give you a better understanding of clock jitter and how it affects the performance of the high-speed ADC. - C0 Q. k a1 H$ d: k Z+ X
$ P' v' Z3 C+ ]% v r+ X- R. hAs an example, we will highlight the latest high-performance ADC from Linear Technology, the 16-bit, 160 Msps LTC2209. This converter exhibits a signal-to-noise ratio (SNR) of 77.4 dB, with 100 dB spur free dynamic range (SFDR) throughout much of the baseband region.
* J& B, x1 P7 G
3 u/ C9 e* ?' F; XLike most high-speed ADCs on the market today, the LTC2209 uses a sample-and-hold (S&H) circuit that essentially takes a snapshot of the ADC input at an instant in time. When the S&H switch is closed, the network at the input of the ADC is connected to the sample capacitor. At the instant the switch is opened, one-half clock cycle later, the voltage on the capacitor is recorded and held.
+ r& b/ y* W3 W1 k4 Y9 x! a
8 P `/ c0 v# f+ C& C) OVariation in the time at which the switch is opened is known as aperture uncertainty, or jitter, and will result in an error voltage that is proportional to the magnitude of the jitter and the input-signal slew rate. In other words, the greater the input frequency and amplitude, the more susceptible you are to jitter on the clock source. Figure 1 demonstrates this relationship of slew rate being proportional to jitter. ( j- w3 }; a7 Y
2 H0 U) p) W3 E* x/ V, |( ?% {
# _8 r) L/ j$ [, Y M
" c% S3 i8 _2 U* p4 F
Figure 1: Slew rate exacerbates the effects of clock jitter. # m, L# A7 M8 x* `1 C- R. w
(Click on image to enlarge)4 S7 I4 K2 m. B$ Y/ A
; _, T' B# ]4 ], z
' w! |: h- f3 i9 DDescribing a clock as "low jitter" has become almost meaningless. This is because it means different things to different interest groups. For a programmable logic vendor, 30 picoseconds, or even 50 psec, is considered low jitter. In contrast, high-performance ADCs need a clock with jitter under1 psec, depending on the input frequency. 7 w- E! U- K5 I4 o
2 O9 u- [( p0 \More precisely, spectral power distribution of the sampled signal is the determining factor, as opposed to simply the highest frequency component, unless a full-scale signal at the upper end of the spectrum is expected. For a simplistic example, a uniform band of power from DC to 1 MHz is 6 dB less sensitive than a single tone, or a narrow band, with equivalent power at 1 MHz.
2 Z z \) M9 Q, T7 G
, b; B0 G7 B( X' V7 F0 Q5 r. bThere are various contributors to jitter in any scenario, extending from the oscillator to any frequency dividers, clock buffers and any noise acquired due to coupling effects, in addition to the internal aperture jitter of the ADC itself.
( L7 s! q" o5 G/ v6 r. I
& B7 Z' @7 f. s- b: AThe internal aperture jitter of the LTC2209 is 70 femtoseconds (fsec). For the level of performance exhibited by the LTC2209 and other members in Linear Technology's high-speed 16-bit family, 0.5 psec jitter (the best available from many oscillator vendor) may produce discernable compromise in SNR for some sampling scenarios. It is not the ADC but the sampling scenario that dictates the required jitter performance.
: m: {) R* ^% r( ^$ y7 e- V # H3 T6 |' m$ {% l& P& }$ @
Any ADC that exhibits 77 dB SNR at 140 MHz input frequency would require the same jitter performance to achieve full data sheet SNR. It is the input frequency, not the clock frequency, which is the determining factor with respect to jitter performance. On the LTC2209, a clock that has 10 psec jitter would cause a loss of only about 0.7 dB SNR at an input frequency of 1 MHz. At 140 MHz, the SNR would degrade to 41.1dB. 1 d7 C6 P7 m$ P: ~+ }7 e
Z9 i% \' z* f9 @, j+ B
Figure 2 demonstrates the effects of clock jitter on the SNR of the LTC2209 as a function of sampled input frequency, with a family of curves of increasing clock jitter ranging from a perfect clock to 100 psec of jitter. At 100 psec, the ADC SNR begins to degrade with input frequencies of only 200 kHz! 6 E2 ]! p4 t3 J
& P9 V% W4 B9 u, ]' Q
* V$ n/ e: x' K8 K7 |2 iFigure 2: Jitter degradation of SNR as a function of input frequency.
* ?" ?- D* k* K(Click on image to enlarge)
; L) E; ]6 c" i' ?* K5 D% t
) }/ j8 U- Z4 r5 G
" |. e4 j+ s6 h4 N$ ~) L, e; ^The theoretical limit on SNR resulting from clock jitter is: / W" h$ S3 v4 P" S0 P% I
/ x! _; b3 H' Q% |* @ h# X& W
: [; _: t7 k6 g0 L7 `# `4 O
, l0 r" Y& i2 Z$ W- L
: F; x$ {; t- T& u! J, G Dwhere fin is the input frequency and σ is the jitter in root mean square (RMS) seconds.
* X4 K3 S( j! c4 a# [
8 a/ m q! y" s; R' KThe jitter-related noise power is proportional to the input power (dBFS). As the input level is raised or decreased, the noise component related to jitter changes accordingly, For example, if we have a -1 dBFS input signal at a 70 MHz IF, sampled by a clock with 1 psec jitter, we can expect an SNR of 68 dBFS. At -5 dBFS, the noise component related to jitter would drop 4 dB to an SNR of 72 dBFS.
5 ~( H+ q- P$ |+ L h+ @
( P$ {2 q- b5 X, _' g# {; xTo calculate the total SNR degradation, we add the jitter-noise power to the published SNR of the ADC, 0 g! V' t2 V4 G2 W1 a9 U: D
: k/ }. E2 o! z. `
$ W9 q. k Y, ~. r
| h, v6 W2 G* I
& d' }. u. v4 I1 L' N) R8 v$ }(Click on image to enlarge)Understanding clock oscillator jitter specifications ! T, F1 k6 Q+ ?. c& L2 Y* N. Q
Clock oscillators are usually specified in terms of spectral density of phase noise in dBc/Hz. An oscillator output can be decomposed into an amplitude term with associated amplitude noise and a frequency term with associated phase noise:
3 V0 {, _7 I/ Y8 s7 s: @, Y $ n3 m% a: n5 D4 M, J4 d
! O- S) h$ z) ?3 F" @ N- |' `
9 b, u& S# j4 I. r5 a3 p/ |) c; B
The spectral density measurements assume the AM component ε(t) of the noise is negligible compared to the phase noise component φ(t). This is a reasonable assumption with any quality frequency source. ( H# T- ?2 }/ O2 L
The spectral density denoted as L(f) is stated as the ratio of the single-sideband phase-noise power in a 1-Hz bandwidth at an offset frequency, also called the Fourier frequency, relative to the carrier power:
$ Y( n0 G2 e/ d% q& p9 Q + b: | _. s5 ^5 Q3 x
% a) b2 J* u* H+ ?: s1 ^: p6 n* o" _1 _: I/ I
(Click on image to enlarge)
8 v/ u2 i" M: o* gJitter is the integral of spectral phase density with respect to frequency between two limits in frequency, and expressed in time: 0 s7 n. d* E$ A) s0 f. Q- E* n3 O
- u9 ]9 Z1 s; A0 @7 Q7 e
! m! \, h6 m. v1 \: _8 w
* L& ]# o( R& p7 X
, V6 P7 y( `* D3 w(Click on image to enlarge)
# ]/ T( E! ^- [: SThe result is frequency independent.
* T7 x1 o7 n* V0 R( U6 y ~' O0 p' Q8 l, a H8 {+ ^
Most oscillators that rate jitter are rated between 12 kHz and 20 MHz. This is due to historical reasons related to optical communications and is not applicable to most other practical cases. Performance may, in fact, fall apart beyond these limits, so take care not to be lured in without careful examination.
$ F, t' Q- N8 b4 D% ?6 y8 h0 F 0 r" R7 e1 ]4 k9 L* W& ]5 j( M; b
For many oscillators where close-in phase noise dominates, the lower limit has the most impact on the published figure. While this expression is convenient, as it yields a single number useful for calculation of ADC SNR degradation, it is not as informative as the spectral density. . T4 @3 @2 M. | D1 j8 v
# \, |, P/ Z, ]2 X# R3 uFor example, two oscillators having different spectral content may have the same jitter over the same integration limits but may not produce the same SNR. Elevated wideband noise may not produce a poor jitter spec, but will degrade SNR. Close-in phase noise causes the fundamental signal to spread into adjacent frequency bins of an FFT, thereby reducing dynamic range, whereas broadband phase noise will uniformly elevate the noise floor throughout the entire Nyquist zone, thus reducing the overall SNR performance of the ADC. + f9 d9 u5 g$ x. G$ t2 f, I
/ X/ D( C# X3 o, s' S+ ]+ y+ z
Jitter does not affect SFDR unless the clock also contains spurs. The lower frequency limit of integration should correspond to the frequency resolution of any manipulations of the sampled data as the size of an FFT increases, for example.
/ s1 o. |: s* r' V4 [' y ( b& J, \9 v" e8 W$ |
Figure 3 shows the effect of band-limited clock jitter related to phase modulation of two signals of similar amplitude, but of different frequencies. This illustrates the exaggerated effect of both random phase noise and phase modulation of the clock in the presence of higher input frequencies. The clock input of the ADC should be regarded as the local oscillator port of the ADC, not a digital control signal. Anything present on the clock, including wideband noise extending to GHz frequencies, will mix with the input signal.
$ b& Q) M8 o3 D! x5 Z- v, B U# Z! Q0 {
' N6 \" b2 K+ x y4 C0 B' ^& U$ o V$ Z0 L
* w1 C: U, F/ i; M! Z
Figure 3: Effect of band-limited clock jitter related to phase modulation of two signals of similar amplitude, but of different frequencies. ; B* O, p; j7 k) L- e
(Click on image to enlarge)+ d" o8 f' x* n9 e3 {- F
( |- I4 V0 }+ b" C* f$ a
, K+ L6 O7 ]4 w j8 G
( Part 2 will look at how the application, not the ADC, determines the clock-jitter need; selecting an oscillator to drive high-speed ADCs; and clock sources and architectures. You can read it by clicking here.) : ?1 I' }6 ?2 D$ V
. `) K0 R+ X/ ?4 e
Conclusion
0 K4 M$ D. N7 M$ oThe impact of jitter on ADC performance is a function of input frequency (slew rate), not sample rate. The choice of clock source will be determined by the application. Don't always believe the clock-source vendors. Test your clock sources with your ADC evaluation board before it is too late. , B/ c/ {. c; P1 R) r/ h1 @
* G/ _$ o2 U: n' x( E* P+ q4 L
Related articles
* Q P- H( L3 [# F About the authors
6 {4 O9 N4 B5 m3 a3 ]+ i4 q* ADerek Redmayne is a mixed-signal applications engineer, and Alison Steer is a product marketing manager, both at Linear Technology Corp, Milpitas, CA.
6 }2 {- x& t2 c: v0 V8 k" |! i: t" j: c" A/ w
Understanding the effect of clock jitter on high-speed ADCs (Part 2 of 2) $ B, Q" D- v M/ P8 l; O; c
Derek Redmayne & Alison Steer, Linear Technology Corp.8/14/2008 12:00 PM EDT(You can read Part 1 here) # l$ c1 j$ K) w: u
The application determines the needs, not the ADC& a U: i- P$ J/ u
Applications which are receiving weak signals which are in close proximity to strong tones, such as.static reflections in Doppler ultrasound, radar, and RFID, are sensitive to close-in phase noise. Conversely, when digitizing a CCD output, jitter generally doesn't matter due to the low slew rate at the point in time at which sampling occurs. Video applications are also not very sensitive. For example, in HDTV the sample window is approximately 6400 psec (time per pixel). * _; Y& x7 E; U+ ]
High symbol-rate communications applications are generally not sensitive to close-in phase noise, and may not be overly sensitive to the effects of wideband phase noise. High crest-factor waveforms (WCDMA OFDM) with relatively even power distribution have a low RMS power level, and also require headroom, so will not elevate the noise floor as much as a full amplitude single tone. However, higher-order modulation types, such as QAM and M-nary phase modulation, are more susceptible to noise and have more narrow carrier-recovery loop bandwidths for the same symbol rates as, for example, QPSK used in CDMA systems. ' c/ V- |# }9 Y; _
A digital radio where strong interferers (single tones) may appear in close proximity, or may be much stronger than the signal of interest, is generally demanding in terms of close-in phase noise, and may be sensitive to wideband phase noise. As any wideband signal source tends to have a high crest factor, and requires headroom for interferers, the nominal power at the ADC may be low. The characteristics of the band of interest must be taken into consideration in deciding on a clock source.
$ G$ \6 u/ L- g# j: D6 F: P4 `0 {+ |Selecting an oscillator to drive high-speed ADCs+ g8 D& K- \) w( W* i6 ^
Most oscillators will have close-in phase noise that will limit the dynamic range close to a strong fundamental. If close-in phase noise is important, based on your dynamic range requirement in proximity to strong tones, you may need a phase locked loop (PLL) to reduce the close-in noise of your oscillator source, or to lock your oscillator to an accurate frequency reference. The use of a PLL as a jitter cleaner essentially provides a very narrowband tracking filter.
1 J5 r* ^- j8 o1 _8 _+ ~Your choice of oscillator will dictate your loop bandwidth, as well; your desired loop bandwidth will dictate the oscillator. A voltage-controlled crystal oscillator (VCXO) requires only a narrow loop bandwidth to track a stable reference. VCOs can provide wide tuning range, but need wider loop bandwidth in order to reduce their close-in phase noise to acceptable levels.
* D. Y, `* v9 t$ V3 T9 R! P: LIf you only require a very restricted tuning range, perhaps locking to a reference oscillator, the use of a VCXO is the best option. If you need the octave tuning range of a VCO, and need low close-in phase noise, you may have a problem, especially if you need high divider ratios and low reference-comparison frequencies in your PLL. Figure 4 shows a real VCXO phase noise plot, compared to a typical VCO. : ]+ M8 Q: L' Z/ j3 }
' [5 |) s$ P5 O z( Q; x# zFigure 4: Comparison of VCXO versus hypothetical VCO phase-noise performance8 A0 n5 F r9 [! [8 h
(Click on image to enlarge)
" [! y& ~# w5 x4 e" x o. \6 w
4 j: |* Y0 `! v/ R+ IThe optimal loop bandwidth for the PLL is suggested by the intersection of the noise density of the reference oscillator as multiplied by the center frequency, and the phase-noise plot of the VCXO or VCO. The example would suggest 2 kHz for the VCXO, and 300 kHz for the VCO. A 300 kHz corner requires a comparison frequency of at least 3 MHz, which suggests 5 MHz. 4 B' E- c! O9 ^9 I7 r. [
The VCXO could be used with a comparison frequency as low as 20 kHz. If a lower frequency reference were used (high divide ratios) with the VCO, the intersection of the multiplied phase noise with that of the VCO would be at a lower frequency, substantially increasing the jitter. The use of an excessively low loop bandwidth with a lower multiplication ratio will cause the phase noise of a VCO to remain within the loop bandwidth. If your application is insensitive to close-in phase noise, and does not need to be locked to a reference, an XO can be used. 7 R* ~* R0 X& d- l0 x. G' m5 ~
Clock sources and clock architectures8 Z0 H, C5 b' I+ [6 H
A good clock can be compromised by routing it through an FPGA where internal crosstalk is prevalent. FPGAs often maximize their input/output connections (I/Os) at the expense of ground pins, resulting in ground bounce. If the FPGA is driving outputs at different rates, these will manifest themselves in any clock routed through the FPGA, and ultimately on the output of any ADC using that clock.
: f$ Y7 U6 i) p' `( x8 t1 K, |( gA low-noise flip-flop clocked by the clean VCO signal can be used as a retiming stage to eliminate jitter when an FPGA is used to frequency-divide the VCO. The FPGA can be used to implement a narrow-band PLL for an external VCXO, with an external loop filter, and a loop-filter driver protected from reflected ground bounce from the FPGA. Do not use a digital lock loop (DLL) to produce a clock for an ADC unless you are over-sampling the audio band. 0 E& y5 p' h8 m# v5 e* N" U
A good clock can also be compromised by routing it among digital signals. Any clock originating at any distance from the ADC must be routed through a conduit of copper and vias. Figure 5 shows examples of good and bad routing of clocks. The bad cases are where the clocks are within cavities shared with digital signals. " U; R1 q1 i/ H K6 T1 r' Q
5 l5 _+ \+ R# n4 g8 g+ }. gFigure 5: Example of good and bad layout for clock routing.
/ d& C9 h3 O8 w2 p(Click on image to enlarge)
1 Z5 a- r$ g0 m0 M
' v4 Y @+ a- ~) c+ t2 c9 U2 RConclusion
' L0 K" I8 [7 Z$ L4 f5 x) ^; k% @The impact of jitter on ADC performance is a function of input frequency (slew rate), not sample rate. The choice of clock source will be determined by the application. Don't always believe the clock-source vendors. Test your clock sources with your ADC evaluation board before it is too late. % I! r' @) m" c, Y, T2 l
Related articles. G- D9 u* v; m8 [9 g% z" [; M
About the authors
- `6 w7 e: \ l/ a& A8 ]Derek Redmayne is a mixed-signal applications engineer, and Alison Steer is a product marketing manager, both at Linear Technology Corp, Milpitas, CA.% c% H" F) O$ m; y
$ |/ K& w* ^8 b
8 ?) e$ u# \( J9 _3 W
. {' d+ `6 j, O" A9 g: R |
|