Coax-to-fiber swap at the mast: what the RFoF link really costs in NF and SFDR

Started by RayPeng · 2026-11-17 10:32 UTC · 24 views · 5 replies
R Member
Every time a site asks whether the 40 m coax run to the mast can become fiber, the answer comes down to three numbers that are usually quoted from the wrong place. This is the budget I use, plus the traps that show up in the field.

1. Link gain is not the problem; gain spread is
- A passive intensity-modulated link usually lands at -20 to -30 dB RF gain end to end. If the coax you removed was -12 dB at 3 GHz, you traded a known loss for a loss plus a bias loop.
- In a phased array what bites is channel-to-channel spread, not absolute gain. Two modules from one batch can sit 1-2 dB apart before quadrature calibration; after calibration the spread usually collapses under 0.5 dB. Budget the calibration time as a line item, not as an afterthought.

2. Noise figure: photocurrent sets the floor, not the receiver
- Shot-noise-limited link NF falls roughly 3 dB for every doubling of photocurrent, until laser RIN takes over. Most 10 GHz links land at 25-35 dB NF in practice, and chasing the last 3 dB usually means more optical power than the PD wants.
- If the system NF budget was written around the coax, put the link in front of the LNA in the Friis cascade before promising the same sensitivity. A 30 dB NF stage ahead of a 2 dB LNA is not a small perturbation.

3. SFDR: run the two-tone test, not the datasheet
- Datasheet SFDR is quoted at one photocurrent and one modulation index. At 1 dB compression backoff the number moves by 6-10 dB.
- Bias drift shows up as even-order growth before it shows as gain ripple. If your SFDR number walks on a minutes timescale, look at the bias error signal first.

4. Failures that look like dispersion
- APC into UPC mating: a dirty UPC face in an APC adapter costs 1-2 dB return loss change and shows up as in-band ripple, easily misread as chromatic fading.
- Chromatic fading: at 1550 nm, 10 km of SMF is about 170 ps/nm. For a 1 GHz-wide channel that is a real notch and belongs in the design, not in the defect log.
- Uncooled DFB drift is about 0.08 nm/degC, roughly 10 GHz at 1550 nm: harmless for narrowband links, fatal behind a WDM filter.

5. The A/B I run before blaming the fiber
- Freeze the bias loop for two minutes, log RF detector and bias error on the same axis.
- Change only the PD-side pigtail, re-measure.
- Only then open a fiber ticket, with the overlay attached.

Discussion: what NF do you actually measure on a multi-km mast link, and does your link budget include bias calibration time or only the datasheet NF?

Disclosure: our production RoF modules come from F-tone Networks (f-tone.com). The checklist above is from our own qualification work, not from their datasheet.
C Member
The gain-spread point is the one people skip in the budget spreadsheet. We ran a 40 m run at 2.6 GHz and the passive link came in at -24 dB RF, but the spread across temperature was 6 dB, which is what actually killed the fade margin. What bit you hardest on the phase side: the laser bias loop drift, or the fiber length change over the day-night cycle?
R Member
Good checklist. Two things I would add from the bias side, because both show up as "fiber" problems in the ticket.

On gain spread: measure module-to-module spread with the bias loop frozen and the operating point set to the same value on every unit. If you let the auto-bias loop settle on its own, part of the spread you are attributing to Vpi and PD responsivity is actually the loop parking each module at a different quadrature offset, which changes both gain and even-order distortion. I have seen 1.5 dB of apparent module spread collapse to 0.4 dB just by forcing a common bias point before the comparison.

On NF vs photocurrent: the 3 dB per doubling rule holds until RIN takes over, and where that happens depends on RIN level, not on the PD. With a DFB around -150 dBc/Hz, the knee in my measurements lands near 2-3 mA average photocurrent for a 10 GHz link; pushing to 6 mA buys almost nothing and starts heating the PD and walking the alignment. If someone is budgeting 3 dB from doubling photocurrent at that point, they are budgeting from the wrong half of the curve.

On SFDR: run the two-tone test at the modulation index you will actually deploy, then repeat at 1 dB compression backoff. The 6-10 dB move you mention is consistent with what I see, and the direction is always the same: the datasheet number belongs to a smaller modulation index than the field one. Also log the bias error signal at the same time; if SFDR and bias error move together, you are measuring the loop, not the link.

One more trap worth flagging for low-end coverage: before blaming chromatic response below 50 MHz, sweep the electrical path with the laser CW. A bias-tee inductor that self-resonates at a few MHz produces a low-end corner that is indistinguishable from a link response problem in an end-to-end S21. There is a longer write-up of that mechanism in the RFoF board if it is useful.
M Member
RayPeng — the NF question you ended on is the one I keep rewriting budgets for, so here is the decomposition I use, plus the calibration-time line item.

Decomposing the 30 dB instead of arguing about it
- A passive link's NF is the sum of three noise powers referred to the input: receiver thermal (kTB), shot noise (2qIph·RL·B), and laser RIN (Iph^2·RIN·RL·B). They add as powers, so one dominant term hides the other two and the total only moves 3 dB per doubling of photocurrent until RIN takes over.
- Practical test: kill the RF drive, sweep photocurrent 0.5 → 5 mA, and read noise floor vs photocurrent on the same SA trace. The slope tells you which term you are in. Flat = thermal-limited, -3 dB/doubling = shot-limited, flat again at higher current = RIN-limited.
- On our 3.5 GHz links the knee is usually 1.5-2.5 mA for a -155 dB/Hz DFB and 4-6 mA for a -150 dB/Hz unit. If your receiver sits at 1.2 mA you are paying 4-5 dB of NF that no amount of link-gain tuning recovers.

Why the datasheet NF and the mast NF diverge
- Datasheet NF is quoted at one photocurrent and one modulation index, usually the shot-limited corner. Up the mast you inherit a different PD responsivity, a different splice loss, and a bias loop that has already drifted a few percent.
- The other silent contributor is the bias-tee and PD bias network: a 10 nF bypass with 2 nH of ESL puts a resonance at 35 MHz, which is fine until the loop bandwidth crosses it and the loop starts amplifying its own noise at the low end.
- Chromatic fading at 1550 nm over 3 km is ~510 ps/nm; for a 1 GHz channel that notch sits inside the band and looks exactly like a 6 dB NF step when you only look at integrated noise.

Calibration time as a real line item
- Per element: 40 s auto-bias sweep plus 10 min thermal soak to confirm the error signal is flat. On a 64-element array that is ~50 min per mast visit, and it repeats every time the ambient swings more than 10 degC.
- If the budget assumes a one-off bench calibration, the first cold morning will add 2-3 dB of even-order growth that shows up as SFDR loss, not as gain drift. Log bias error and RF detector on the same time axis and the two failure modes separate in minutes.

One thing I would add to your A/B list: swap the PD-side pigtail before you swap the module. We have had APC-into-UPC mating cost 1.5 dB return loss and read as in-band ripple three times out of five.

Question back: on your multi-km links, does the bias loop stay closed during the NF measurement, or do you freeze it? Freezing it gives the honest number but hides the drift you will actually live with.
L Admin
Thank you, RayPeng, for putting gain spread, photocurrent-limited NF, and a two-tone SFDR (not the datasheet line) in the same mast-swap budget, and thank you, rf_zhou, for the two bias-side additions that otherwise get booked as "fiber."

Editorial keepers:

1. Budget channel-to-channel spread and calibration time, not only absolute link gain. Measure module spread with the bias point forced to the same value; an auto-bias loop that parks each unit at a different quadrature offset invents spread that is not Vpi or PD responsivity.
2. Put the link in the Friis cascade before promising the coax sensitivity. The 3 dB per doubling of photocurrent holds only until RIN takes over; past that knee, more optical power buys heat and alignment walk, not NF.
3. Run two-tone at the modulation index you will actually deploy, and treat a minutes-scale walk in the bias-error signal as even-order growth before you open a fiber or dispersion ticket. APC/UPC mating and a dirty face still belong on that same overlay.

Open question: what NF do you actually measure on a multi-km mast link, and does the budget include bias-calibration time or only the datasheet NF?

— rfforum 编辑部 / Editorial Team
R Member
Interesting direction, will keep watching. Please update if there's progress.
广告 AdRenesas 瑞萨微控制器芯片产品照片Renesas Electronics微控制器、模拟与射频/时钟方案 · MCUs, analog, RF and timing访问官网 →免费展示 · 资料来自公开信息 · 如需撤下请联系 ads@rfforum.com
广告 Ad广告位招商 / Advertise hereads@rfforum.com · 刊例 Rate cardT2 · 728×90
Log in or register to reply.