Published Nov 20, 2026 · Last updated Nov 20, 2026 · 5 min · IndieRF
ADC Clock Jitter from Phase Noise: 94.7 fs
LMK04828 into an ADS42LB69: a 79.9 dB SNR limit at 170 MHz.
The LMK04828's printed phase noise at 245.76 MHz is 94.7 fs rms from 12 kHz to 20 MHz, a 79.9 dB SNR limit at a 170 MHz input. The ADS42LB69 prints 73.2 dB there, so the ADC sets the SNR and the clock costs 0.8 dB.
Six printed points, integrated as straight lines
TI prints the LMK04828's closed-loop phase noise at six offsets for the HSDS 8 mA output, VCO0, with a Crystek CVHD-950 VCXO on PLL1: −124.3 dBc/Hz at 1 kHz, −134.7 at 10 kHz, −136.5 at 100 kHz, −148.4 at 1 MHz, −159.1 at 10 MHz and a −159.3 floor at 20 MHz. Plotted on log-log axes those are straight lines, and a straight line there is a power law, . Each segment integrates exactly:
with at , the −10 dB per decade case. Double the sum for both sidebands and take the root: 0.146 mrad, or 0.00838°, rms. Divide by at 245.76 MHz and that is 94.7 fs.
The segment from 100 kHz to 1 MHz carries 39% of the phase variance. That is where the PLL2 loop bandwidth sits, 344 kHz for VCO0 in TI's test setup, and it is where a measured trace and straight lines through six points part ways. TI prints 99 fs measured over the same band; the straight lines land 4% under that, about 0.4 dB of SNR limit.
The lower edge of the band matters less than people expect here. From 1 kHz the same mask gives 98.6 fs, and from 100 kHz it gives 84.2 fs. TI also prints 102 fs from 100 Hz, which the calculator refuses to reproduce: the table starts at 1 kHz, and a level below the first printed point is a guess.
What a spreadsheet trapezoid gives from the same table
The usual sheet takes the antilog of each dBc/Hz value and runs the trapezoid rule between points. That is SiTime's published method, and it asks for at least 20 points per decade so the chords follow the curve. A data sheet table gives one per decade.
Through the printed points from 12 kHz, the trapezoid reads 134 fs, 41% high. On a slope falling 10 dB per decade the straight chord in linear units sits far above the power law over most of the decade, and every segment here falls. In SNR limit that is 3.0 dB at every input frequency: 76.9 dB instead of 79.9 dB at 170 MHz. That is enough to send someone shopping for a better clock they do not need. Read 20 points per decade off the same straight lines and the trapezoid comes back to 94.8 fs.
Which part sets the SNR at each input frequency
Clock-limited SNR is , so it falls 6 dB per doubling of input frequency, and the ADC's own SNR barely moves. The ADS42LB69 numbers are typical, 2 Vpp full scale, −1 dBFS, measured at 250 MSPS; at 245.76 MSPS the Nyquist band is 2% narrower, under 0.1 dB. The 170 and 230 MHz rows sit in the second Nyquist zone, which is the IF-sampling case this ADC is built for.
| ADC input (MHz) | ADS42LB69 SNR, printed (dBFS) | Clock limit, 94.7 fs (dB) | Both, power sum (dB) | Clock limit, trapezoid 134 fs (dB) |
|---|---|---|---|---|
| 10 | 73.9 | 104.5 | 73.9 | 101.5 |
| 70 | 73.7 | 87.6 | 73.5 | 84.6 |
| 170 | 73.2 | 79.9 | 72.4 | 76.9 |
| 230 | 72.8 | 77.3 | 71.5 | 74.3 |
Rule of thumb: a clock limit 10 dB above the ADC's own SNR costs 0.4 dB, and 6 dB above costs 1.0 dB. At 170 MHz the margin is 6.7 dB and the cost 0.8 dB. The power sum reads slightly pessimistic, since the printed SNR already carries TI's bench clock and the ADC's 85 fs aperture jitter.
I'd keep this clock for inputs up to about 230 MHz, where it costs 1.3 dB. Past that the limit drops 6 dB per octave while the ADC's SNR holds near 73 dB, and the clock takes over.
The same clock in the receiver lineup
In Cascade, a mask on an LO block prints phase error and jitter rows down the chain, and the ADC block takes the clock's rms jitter and prints the clock-jitter SNR limit beside its effective SNR. The linked design is a 170 MHz tone at −1 dBFS into the ADS42LB69 at 245.76 MSPS, with 94.7 fs on the ADC: effective SNR 72.2 dB (the printed 73.2 dBFS, 1 dB below full scale) and, in the same column under Phase noise and jitter, the clock limit at 79.9 dB. Cascade flags the 1 dB clipping margin, which is the level TI measures at. Full scale is set to +10 dBm, 2 Vpp across a 50 Ω termination, which is a board choice; the ADC field holds the clock alone because the printed SNR already includes the aperture jitter. Move the tone to 230 MHz and the limit row follows.
References
- Texas Instruments, LMK0482x Ultra Low-Noise JESD204B Compliant Clock Jitter Cleaner With Dual Loop PLLs, SNAS605AS, March 2013, revised May 2020, section 7.5 Electrical Characteristics: page 17 (noise floor at 20 MHz offset), page 18 (closed-loop phase noise, LMK04828 VCO0, 245.76 MHz), page 19 (closed-loop jitter), notes (15) and (16).
- Texas Instruments, ADS42LB49, ADS42LB69 Dual-Channel, 14- and 16-Bit, 250-MSPS Data Converters, SLAS904F, October 2012, revised May 2016, section 6.5 (SNR, ADS42LB69) and section 6.9 (aperture jitter).
- SiTime, Phase Noise Spreadsheet Calculator, the trapezoid method and its 20 points per decade.
Related
Questions? Contact
Discussion
No comments yet. Start the thread with a measurement, a correction, or a worked example.