Almost every radiated emissions failure in a digital product traces to the same source: a clock edge, repeated at a fixed frequency, with harmonics that march up through the spectrum and land on the measurement band. The engineer then faces a choice that looks binary. Slow the edge or modulate the clock and reduce the emission at birth, or keep the signal clean and put a metal can over it.
In practice neither technique is universal. Spread-spectrum clock generation is nearly free and can recover 6 to 12 dB on selected harmonics, but it does nothing for the fundamental, it can violate interface specifications, and it fails outright when the spread is narrower than the measurement bandwidth. Shielding delivers 20 to 40 dB but costs money, adds rework, and hides problems rather than solving them. This guide quantifies both so the choice stops being a matter of taste.
How Spread-Spectrum Clocking Works
Instead of a fixed frequency, the clock generator slowly varies its output, typically in a triangular or Hershey-kiss profile, over a band of 0.5 to 2.5 percent of centre frequency at a modulation rate of 30 to 33 kHz. Down-spread modulation shifts the frequency lower and back; centre-spread moves symmetrically. The energy that would have concentrated in a narrow carrier is redistributed across the modulation band.
The bandwidth equation that decides everything
Emissions measurements use a resolution bandwidth of 120 kHz between 30 MHz and 1 GHz, and 1 MHz above that. Peak reduction is therefore limited by how the spread compares with that window. If the total frequency excursion is much larger than the resolution bandwidth, the reduction approaches ten times the logarithm of the ratio of bandwidth to excursion. If the excursion is smaller than the window, the receiver still integrates the whole spread and the reduction is close to zero.
A 100 MHz clock with one percent down-spread sweeps 1 MHz. Against a 120 kHz window that ratio is 8.3, so the theoretical improvement is about 9 dB. A 25 MHz clock with the same percentage sweeps only 250 kHz, giving roughly 3 dB. The same setting that rescues a 100 MHz processor bus does almost nothing for a 25 MHz interface clock. This is why SSCG results vary so much between projects and why vendors who quote a flat number are usually quoting their best case.
Harmonic order changes the answer again
Spread is a fixed percentage of the fundamental, so the absolute excursion grows with harmonic number. The tenth harmonic of a 25 MHz clock sits at 250 MHz with a 2.5 MHz excursion and benefits fully. The fundamental sits at 25 MHz with 250 kHz and benefits barely at all. Expect roughly 2 to 4 dB at the fundamental, 5 to 8 dB by the fifth harmonic, and 9 to 12 dB beyond the tenth, with diminishing returns once the spread exceeds several times the bandwidth.
Where Spread-Spectrum Clocking Fails
The technique is not free of consequences, and the failure modes are contractual rather than physical.
- Synchronous interfaces that recover a clock assume a stable reference. PCI Express defines its own SSC rules and limits; feeding it a clock with the wrong modulation profile or depth produces link training failures that appear as intermittent drops, not as obvious errors.
- USB 3, SATA, DisplayPort and MIPI D-PHY all specify reference clock jitter and modulation limits. Exceeding them may pass EMC and fail interoperability.
- RF and conversion subsystems care about phase noise, not average frequency. Spreading the sampling clock raises the noise floor near the carrier and can degrade signal to noise ratio by a fraction of a bit.
- Some automotive and avionics specifications treat intentional modulation as a design change requiring re-qualification.
- Any product where the clock is distributed to a customer interface may be contractually prohibited from spreading it.
The practical rule is to spread only what is internal. Modulate the processor and memory clocks if the memory controller tolerates it, leave the interface reference clean and shield that one instead.
What Shielding Actually Buys
A well grounded shield can on a 20 by 20 mm module typically delivers 30 to 45 dB between 300 MHz and 3 GHz, provided three conditions hold: the can is bonded to a continuous ground plane, seam spacing stays below one twentieth of a wavelength at the highest frequency of interest, and no cable or trace crosses the boundary without filtering. Break any one of them and the measured performance drops to 10 or 15 dB, which is the number most disappointing shielding measurements actually show.
Cost is the other axis. A two piece shield adds roughly 0.15 to 0.40 US dollars in volume, plus placement, plus rework risk, plus the board area lost to a keep-out for the ante
a. Compared with a register write that enables spread spectrum, the economics are not close. Compared with a redesign of a return path that is radiating because a plane was split, shielding is also not the answer.
The Correct Order of Attack
Source suppression beats containment every time, and spread spectrum is not the first source technique to reach for.
- Fix the return path first. A signal that returns on a split plane radiates no matter what covers it. Check that every high speed trace has an uninterrupted reference, and that plane splits are bridged with stitching vias at intervals under one twentieth of a wavelength.
- Reduce edge rate at the driver. Series resistors of 22 to 47 ohms on clock lines cut high harmonic content with less impact on timing than any modulation scheme, at the cost of a little rise time budget.
- Filter at the boundary. Common mode chokes and feedthrough capacitors on every cable leaving the enclosure convert a 20 dB shielding problem into a 3 dB one.
- Then consider spread spectrum where the interface permits it, and where the harmonic that fails is far enough above the fundamental for the bandwidth equation to work in your favour.
- Shield last, and shield specifically: the culprit module, not the whole board.
Measuring the Benefit Before You Commit
Near-field probing is enough to rank candidates. Scan the board with a 5 mm H-field loop co
ected to a spectrum analyser in peak-hold, with a 120 kHz resolution bandwidth and the same span you will use in the chamber, and record the harmonic amplitudes with spread spectrum on and off. If the peak drops by the predicted amount, the mechanism is working. If it does not move, the emission is not coming from the clock you modulated, and further register writes will not help.
Then verify the interface. Run the link at temperature extremes and count errors over at least an hour per condition. Spread spectrum failures are statistical and love corner conditions. A marginal eye at 25 degrees C often becomes a link drop at 70 degrees C, and by then the schematic is frozen.
Both techniques belong in the toolbox. The mistake is treating spread spectrum as a free win and shielding as a defeat. One is a targeted 8 dB with side effects, the other is a 35 dB blunt instrument with a bill. Knowing which harmonic is failing and why is worth more than either.