Note 27Signal integrity Rev 1 Watch the video · 0:10

Car camera coax

How does one coax carry a car camera’s video, control and power?

Power rides along as DC, and the data rides on top of it in both directions at once. Four coils keep the data out of the 12 V supply, and each chip subtracts its own signal to hear the other end. On Antmicro’s GMSL2 boards the coils hold 1 kΩ or more over 86% of the band ADI recommends, and over 15 m of coax the video loses 15.9 dB at 3 GHz, where ADI allows 21.1 dB.

The camera sends 6 Gb/s of video and its status replies. The computer sends back control data at 187.5 Mb/s, and 12 V to run the camera. The shield is the ground for all of it.

The figure is 20 ns of the voltage at the camera’s end of the coax.

Four jobs, one coax

20 ns at the camera’s connector

The gray trace is everything on the connector: 12 V, the video going out and the control data coming in. Take away the camera’s own signal, low-pass the rest, and the bold trace is what’s left: the control data.

The numbers

Every 1.7 ns of the 20 ns shown. The control data alone is what arrives from the computer, before the camera’s own signal is added.
Time	Video alone	Everything	Canceled and low-passed	Control data alone
0.0 ns	12.200 V	12.105 V	11.901 V	11.890 V
1.7 ns	11.800 V	11.702 V	11.896 V	11.888 V
3.3 ns	11.800 V	11.684 V	11.898 V	11.898 V
5.0 ns	11.800 V	11.740 V	11.913 V	11.958 V
6.7 ns	12.056 V	12.092 V	11.969 V	12.051 V
8.3 ns	12.056 V	12.138 V	12.048 V	12.096 V
10.0 ns	11.944 V	12.069 V	12.093 V	12.107 V
11.7 ns	12.056 V	12.179 V	12.106 V	12.110 V
13.3 ns	11.944 V	12.039 V	12.108 V	12.108 V
15.0 ns	11.944 V	12.032 V	12.106 V	12.073 V
16.7 ns	12.056 V	12.056 V	12.064 V	11.982 V
18.3 ns	11.800 V	11.731 V	11.984 V	11.916 V

Two boards and a coax

Antmicro’s open GMSL2 pair

Antmicro publishes both ends of a GMSL2 link as open hardware. The camera end is a 43 mm square board with a MAX96717 serializer. The camera’s MIPI CSI-2 lanes go in, and one coax comes out.

The computer end is 75 by 60 mm, with a MAX96724 deserializer that takes four cameras. Both boards use FAKRA connectors and 50 Ω coax.

On each board the coax’s center pin meets two paths. One goes through a 100 nF capacitor to the chip. The other turns off at a tap into a ladder of four coils that ends at 12 V.

Camera end and computer end

Data
12 V

Both boards to one scale, from Antmicro’s renders. The copper from each coax pin is drawn from the board files. From the pin to the tap it carries both. The coax is drawn short.

The numbers

Board	Antmicro design	Chip	Coax connectors
Camera end	GMSL Serializer rev 1.1.1, 43 × 43 mm, 4 layers	MAX96717	J2, one FAKRA
Computer end	GMSL Deserializer rev 1.0.3, 75.1 × 60.1 mm, 4 layers	MAX96724	J2 to J5, four FAKRA

Drawn as a circuit, each end is a bias tee. The capacitor passes the data and blocks the DC. The coils pass the DC and block the data.

The two ends are built the same, part for part. Their coils are filter 14 in ADI’s list of power filters recommended for new GMSL2 designs, and Coilcraft’s reference filter SMD-POC-006.

The computer end puts 12 V onto its coax. The camera end takes it off and feeds its regulators.

Both bias tees

Standard symbols, with the boards’ part numbers. The ESD parts on each line and the 12 V decoupling are left out. The table lists them.

The numbers

Part	Camera end	Computer end
Coupling capacitor	C30, 100 nF	C27, 100 nF
470 nH, Coilcraft PFL1609-471	L6, L5	L13, L9
6.8 µH, Coilcraft 1210POC-682	L4, with R17 across	L5, with R24 across
22 µH, Coilcraft MSS6132T-223	L3, with R16 across	L1, with R19 across
Damping resistors	R17, R16, 5.1 kΩ	R24, R19, 5.1 kΩ
12 V decoupling	C10 100 nF, C11 10 nF	C6 100 nF, C10 10 nF
ESD protection	D2, between the tap and C30	D4, at the chip

Power rides as DC

One coil isn’t enough

A coil’s impedance grows with frequency, so it passes DC and blocks fast signals:

|Z|=2πfL

=2π×10 MHz×22 µH=1,382 Ω

That should make a 22 µH coil a wall at 10 MHz, and Coilcraft’s model of the real part gives 1,451 Ω there. But every coil has some capacitance across its winding. At its self-resonant frequency the two cancel:

f_SRF=12πLC_p

=12π19.7 µH×1.72 pF=27.3 MHz

The model’s inductance sags a little as frequency rises, to the value above at resonance. Past that point the coil acts like a capacitor, and its impedance falls. By 3 GHz the 22 µH coil is down to 31 Ω.

One 22 µH coil

Coil
2πfL
1 kΩ band

The board’s 22 µH coil against a perfect one. The black line is UG-2212’s recommended 1 kΩ, drawn over its band, 2 MHz to 3 GHz. The circle is the self-resonance.

The numbers

Coilcraft’s model of the coil, fitted to its curve for the whole filter. No DC current through it.
Frequency	Ideal, 2πfL	22 µH coil
1 MHz	138 Ω	132 Ω
2 MHz	276 Ω	262 Ω
10 MHz	1,382 Ω	1,451 Ω
30 MHz	4,147 Ω	13,711 Ω
100 MHz	13,820 Ω	999 Ω
300 MHz	41,471 Ω	310 Ω
1 GHz	138,200 Ω	92 Ω
3 GHz	414,706 Ω	31 Ω
10 GHz	1,382,000 Ω	9 Ω

518

One big coil

1 kΩ over 36% of the band

Field marks. The 22 µH coil holds 1 kΩ or more from 7.3 MHz to 100 MHz. That is 36% of 2 MHz to 3 GHz, on a log scale. It shows 262 Ω at 2 MHz and 31 Ω at 3 GHz.

Look-alike. A bigger coil. More inductance looks like more blocking, but past 27.3 MHz the 22 µH coil is a capacitor. At 3 GHz it shows 31 Ω, where a 470 nH coil still shows 439 Ω.

Range. Self-resonance falls as inductance grows: 676 MHz for the 470 nH coil, 119 MHz for the 6.8 µH and 27.3 MHz for the 22 µH. Coilcraft’s data sheet gives the 22 µH part 21 MHz, typical.

Habitat. Every power filter on a camera coax. ADI’s guide has dropped ferrite beads for the job: “Analog Devices is no longer recommending ferrite beads for PoC designs due to their performance sensitivity to increases in temperature and current.”

519

Four coils, smallest nearest the line

Each strong where the others are weak

No one coil covers 2 MHz to 3 GHz, so the board stacks three kinds. Each peaks at its own self-resonance: the 22 µH coil at 27.3 MHz, the 6.8 µH at 119 MHz and the 470 nH at 676 MHz.

Below its peak each acts like a coil. Above it, each acts like a capacitor.

ADI’s data sheet asks for the smallest coil, the one with the highest self-resonance, on the signal trace. The board puts both 470 nH coils there, then the 6.8 µH, then the 22 µH at the 12 V end.

Three coils, each alone

22 µH
6.8 µH
470 nH

Coilcraft’s models of the three parts, each alone with nothing across it. The circle on each is its self-resonance.

The numbers

Coilcraft’s models of the three parts on the board, each coil alone, no resistor across it.
Frequency	22 µH	6.8 µH	470 nH
1 MHz	132 Ω	48 Ω	3 Ω
2 MHz	262 Ω	93 Ω	6 Ω
10 MHz	1,451 Ω	442 Ω	26 Ω
30 MHz	13,711 Ω	1,307 Ω	75 Ω
100 MHz	999 Ω	6,234 Ω	234 Ω
300 MHz	310 Ω	2,317 Ω	687 Ω
1 GHz	92 Ω	603 Ω	1,228 Ω
3 GHz	31 Ω	201 Ω	439 Ω
10 GHz	9 Ω	67 Ω	195 Ω
Self-resonance	27.3 MHz	119 MHz	676 MHz
Share of 2 MHz to 3 GHz at 1 kΩ or more	36%	45%	14%

In series their impedances add, as complex numbers:

Z_PoC=Z_L6+Z_L5+(Z_L4∥R)+(Z_L3∥R)

=(0.5 + j5.5)+(0.5 + j5.5)+(10.7 + j92.5)+(26.7 + j259.6)=38.4 + j363.1 Ω→365 Ω

That is 2 MHz, the bottom of ADI’s band. The two 470 nH coils add almost nothing there, and the sum stops at 365 Ω. At 3 GHz the big coils have turned into small capacitors, and the two 470 nH coils supply most of the 1,098 Ω.

The 5.1 kΩ resistors across the two big coils flatten their peaks. The 6.8 µH coil does more than fill a gap. Take it out of a filter with one 470 nH coil and the two left resonate together, dipping to 251 Ω at 197 MHz.

The board’s filter

Four coils
Coilcraft
22 µH alone

The whole filter as seen from the tap, against Coilcraft’s published curve for the same four parts. The black line is the 1 kΩ band. The circle is 2 MHz.

The numbers

Coilcraft’s curve is read off its SMD-POC-006 sheet. The model is fitted to it, 8% rms.
Frequency	Four coils	Coilcraft’s curve	22 µH alone
1 MHz	185 Ω		132 Ω
2 MHz	365 Ω	364 Ω	262 Ω
10 MHz	1,834 Ω	1,854 Ω	1,451 Ω
30 MHz	4,556 Ω	4,474 Ω	13,711 Ω
100 MHz	3,207 Ω	3,245 Ω	999 Ω
300 MHz	2,184 Ω	2,294 Ω	310 Ω
1 GHz	2,988 Ω	3,059 Ω	92 Ω
3 GHz	1,098 Ω	1,227 Ω	31 Ω
10 GHz	465 Ω		9 Ω

520

The four-coil filter

1 kΩ over 86% of the band

Field marks. From the tap the filter holds 1 kΩ or more from 5.6 MHz to past 3 GHz. That is 86% of the band, against 36% for the 22 µH coil alone. It shows 365 Ω at 2 MHz and 1,098 Ω at 3 GHz.

Look-alike. The 1 kΩ rule. ADI’s guide recommends 1 kΩ from 2 MHz to 3 GHz and lists this exact filter for new designs. Coilcraft’s own curve for it shows 364 Ω at 2 MHz and reaches 1 kΩ at 5.6 MHz. ADI’s echo limit is loose down there. At 2 MHz it asks for the echo to be only 7.8 dB down, and the whole link’s echo is 23.1 dB down.

Range. Coilcraft’s curve holds 1 kΩ from 5.6 MHz to 3.6 GHz. The model here follows it to 8% rms. Coilcraft rates the filter for 8 MHz to 3 GHz at 0.6 A, and ADI rates it for 1,000 mA at 85 °C and 600 mA at 125 °C.

Habitat. Both of Antmicro’s boards, on the camera’s line and on each of the computer end’s four. On both, the tap sits about 12 mm from the connector.

521

Control data comes back on the same wire

Both ways at once

The video flows out of the camera at 6 Gb/s. The control data flows in at 187.5 Mb/s, on the same center conductor, at the same time.

ADI’s channel spec plans the control data from 2 MHz to 400 MHz and the video from 100 MHz to 3.5 GHz. It says so plainly: “The energy of the forward and reverse channel overlaps in frequency.” So a filter can’t pull them apart. At the camera’s pin, its own signal is about twice the size of the control data arriving.

Both GMSL2 chips carry “an echo cancellation circuit” for this. Each one knows what it is sending. ADI publishes nothing on how its canceler is built. A textbook canceler learns how its own signal looks at the pin, reflections included, from the bits it sent. It subtracts that copy, and the other end’s signal is what’s left.

The control data’s eye at the camera

After the camera’s low-pass filter, two control bits wide. Alone has no video on the wire. Low-pass is everything on the pin, filtered. Canceled is the pin minus the camera’s learned copy of its own signal, then filtered.

The numbers

Over the last 1200 control bits of 2400, sampled where the control data alone opens widest.
At the camera’s receiver	Eye opening
Control data alone	161 mV
Everything, low-passed	Shut, 77 mV overlap
Own signal subtracted, then low-passed	142 mV

522

Echo canceling

142 mV of 161 mV

Field marks. With no video on the wire, the control data’s eye opens 161 mV. With the video on and only the low-pass to remove it, the eye shuts. Subtract the learned copy first, and it opens 142 mV, 89% of the clean eye.

Look-alike. A low-pass filter. The control data is slow, so filtering out the fast video looks like enough. In the model the low-pass alone leaves the eye shut. ADI’s channel spec gives the reason: the two channels’ energy overlaps in frequency.

Range. After 12 µs of learning, the camera’s own signal left in the control band is 26 dB below where it started.

Habitat. GMSL2 and GMSL3, at both ends: ADI’s GMSL3 serializer data sheet repeats the same sentence. TI’s training says FPD-Link uses “a proprietary echo cancelation technique” to send both ways over one conductor.

523

Three ways to share a wire

Split, cancel or take turns

Canceling is one of three ways links send control data back. The oldest splits the wire by frequency. TI’s DS90UB913A and 914A keep their control channel near 1 to 5 MHz and their video near 70 to 700 MHz, “because they occupy different spaces in the frequency domain.”

The price is a slow back channel. ADI’s GMSL1 serializer, the MAX96705, sends control data as “a series of 1μs pulses, with a maximum raw data rate of 1Mbps.” ADI’s guide asks GMSL1 power filters to be checked from 50 kHz, not 2 MHz.

The third way takes turns. ASA Motion Link’s physical layer “is based on Time Division Duplexing,” and “Downstream and Upstream traffic does not overlap on the wire.” Each 27.38 µs cycle sends video for 25.92 µs and listens for 0.86 µs, which leaves 0.60 µs for gaps.

Nothing overlaps, so there is nothing to cancel. The video has to make up the time it loses, so the line runs faster than the video:

r_liner_video=T_cycleT_down

=27.38 µs25.92 µs=1.056

That is at least 5.6% faster. Each end also holds the video that waits while the link listens. For 8 Gb/s of video:

B=r_video(T_cycle−T_down)

=8 Gb/s×1.46 µs=11,680 bits=1.46 kB

MIPI publishes no method for A-PHY. A Valens patent, which doesn’t name A-PHY, describes coding the downlink so its low frequencies stay quiet, which “enables band separation between downlink and uplink channels.”

TI publishes nothing on how FPD-Link IV shares its wire. For the DS90UB953, an FPD-Link III part, TI gives the video 20 MHz to 2.1 GHz and the back channel 10 to 50 MHz. The two overlap from 20 to 50 MHz.

Back channels, link by link

Fig. 27g

Link	Video	Back	Sharing
GMSL2	3 or 6 Gb/s	187.5 Mb/s	Both at once, each end cancels its echo
FPD-Link III, 913A	About 70 to 700 MHz	About 1 to 5 MHz	Split by frequency
FPD-Link III, 953	4.16 Gb/s	50 Mb/s	Bands overlap, TI names echo canceling
FPD-Link IV, 971	7.55 Gb/s	47.1875 Mb/s	Not published
MIPI A-PHY 1.1	Up to 16 Gb/s	200 Mb/s	Not published by MIPI
ASA Motion Link	2 to 16 Gb/s	50 or 100 Mb/s	Takes turns

From each maker’s data sheet, training or slides. FPD-Link III’s 913A row gives TI’s bands, not rates. The 953’s overlap is worked from its two bands.

The whole 15 m link

Pin to pin

ADI’s channel spec limits the link from pin to pin: camera board, connectors, coax and computer board together. At 6 Gb/s its equation lets the signal lose up to 21.1 dB at 3 GHz.

In decibels, the losses add along the link:

L_link=L_cam+αℓ+L_comp

=0.87 dB+0.9 dB/m×15 m+1.57 dB=15.94 dB

That is the four coils at each end and new coax, which loses 0.9 dB/m at 3 GHz by ADI’s data sheet. The coils count twice. A weak filter loads the line, which costs loss, and it reflects. ADI limits the echo because the spec uses it to judge “the high-speed transmitter energy reflected into the low-speed receiver.”

The whole link

Four coils
22 µH alone
ADI’s limit

Camera board, 15 m of new coax and computer board, pin to pin, with the same filter at both ends, against the GMSL2 channel spec’s limits for 6 Gb/s. Echoes between the two ends make both curves ripple every 7 MHz. Loss is drawn through the middle of its ripple, and echo along its peaks.

The numbers

ADI’s loss limit is 21.1 dB at 3 GHz. Aged cable is 1.1 dB/m at 3 GHz against 0.9 new. A plus sign is room to spare.
Link	Loss at 3 GHz	Closest to the loss limit	Echo against its limit
Four coils, new cable	15.9 dB	+3.1 dB	+11.5 dB
22 µH alone, new cable	20.6 dB	+0.2 dB	Over from 576 MHz
Four coils, aged cable	18.9 dB	+2.0 dB	+12.2 dB
22 µH alone, aged cable	23.7 dB	−3.0 dB	Over from 735 MHz
Cable alone, new	13.5 dB
Cable alone, aged	16.5 dB

524

The link

15.9 dB of 21.1 dB

Field marks. With four coils at each end, the video arrives 15.9 dB down at 3 GHz, 5.2 dB inside ADI’s limit. The camera board takes 0.87 dB, the coax 13.5 dB and the computer board 1.57 dB. The echo stays at least 11.5 dB under its limit.

Look-alike. Passing on loss. With one 22 µH coil at each end the loss still passes, 20.6 dB against 21.1 dB. The echo breaks ADI’s limit from 576 MHz up.

Range. Aged coax loses 1.1 dB/m at 3 GHz, so 15 m costs 16.5 dB. The four-coil link is then 18.9 dB down and still passes. The lone-coil link is 23.7 dB down, 2.6 dB past the limit.

Habitat. Long runs at 6 Gb/s. ADI’s MAX96792A data sheet rates GMSL2’s 6 Gb/s mode for 15 m of foam coax at 105 °C.

525

Which way back?

From the chips’ part numbers

The serializer and deserializer settle it. Their maker’s papers say how control data gets back, or say nothing.

Key to back channels

How does my camera’s link send control data back?

1aThe chips are ADI’s GMSL partsGo to 2
1bAnother familyGo to 3
2aGMSL2 or GMSL3, such as the MAX96717Both ways at once, each end cancels its echo
2bGMSL1, such as the MAX96705A slow control channel, 1 Mb/s at most
3aTI’s FPD-Link IIIGo to 4
3bAnother familyGo to 5
4aDS90UB913A or 914ASplit by frequency
4bDS90UB953Overlapping bands, and TI names echo canceling
5aASA Motion LinkTakes turns
5bTI’s FPD-Link IV or MIPI A-PHYNot published

Antmicro’s boards end at 2a. For A-PHY, a Valens patent describes separate bands, but MIPI doesn’t say A-PHY uses it.

Stack the coils yourself

Twelve filters, one 22 µH coil

Choose which coils go between the line and the 22 µH coil, and whether 5.1 kΩ resistors sit across the big ones. The curve is the filter seen from the tap.

The readings run the whole 15 m link with that filter at both ends, as in Fig. 27h.

470 nH coils

6.8 µH coil

5.1 kΩ across

The filter from the tap

Passes ADI’s loss and echo limits

Filter
1 kΩ band

At 1 kΩ or more: 86%
Lowest in the band: 365 Ω at 2 MHz
Link loss at 3 GHz: 15.9 dB
Echo: Inside the limit

Coilcraft’s models, as in Fig. 27e. The circle is the filter’s lowest point between 2 MHz and 3 GHz. Loss and echo are judged against ADI’s limits, with new coax.

The numbers

Every filter has the 22 µH coil. The whole link uses the same filter at both ends and new cable.
470 nH coils, 6.8 µH, resistors	At 1 kΩ or more	Lowest in the band	Loss at 3 GHz	Echo
Two, with 6.8 µH, no 5.1 kΩ	86%	366 Ω at 2 MHz	15.9 dB	Inside
Two, with 6.8 µH, with 5.1 kΩ	86%	365 Ω at 2 MHz	15.9 dB	Inside
One, with 6.8 µH, no 5.1 kΩ	79%	361 Ω at 2 MHz	16.0 dB	Inside
One, with 6.8 µH, with 5.1 kΩ	79%	360 Ω at 2 MHz	16.0 dB	Inside
None, with 6.8 µH, no 5.1 kΩ	66%	231 Ω at 3 GHz	16.2 dB	Inside
None, with 6.8 µH, with 5.1 kΩ	65%	230 Ω at 3 GHz	16.3 dB	Inside
Two, no 6.8 µH, no 5.1 kΩ	64%	273 Ω at 2 MHz	16.0 dB	Inside
Two, no 6.8 µH, with 5.1 kΩ	64%	272 Ω at 2 MHz	16.0 dB	Inside
One, no 6.8 µH, no 5.1 kΩ	48%	251 Ω at 200 MHz	16.2 dB	Over from 165 MHz
One, no 6.8 µH, with 5.1 kΩ	48%	267 Ω at 2 MHz	16.2 dB	Over from 192 MHz
None, no 6.8 µH, no 5.1 kΩ	36%	31 Ω at 3 GHz	20.6 dB	Over from 576 MHz
None, no 6.8 µH, with 5.1 kΩ	35%	31 Ω at 3 GHz	20.7 dB	Over from 576 MHz

Conditions

Boards: Antmicro’s GMSL Serializer rev 1.1.1 and GMSL Deserializer rev 1.0.3, Apache-2.0, from their KiCad files: the camera end and channel A of the computer end. The renders are Antmicro’s, with makers’ marks painted out.
Coils: Coilcraft’s lumped models of the PFL1609-471, 1210POC-682 and MSS6132-223. The MSS6132T-223 on the boards has no published model, so the MSS6132-223 stands in, with its capacitance, one resistor and one loss term refitted to Coilcraft’s SMD-POC-006 curve. The PFL1609 model stops at 1 GHz, so 0.07 pF across it was fitted above that. The whole filter matches the curve to 8% rms from 1 MHz to 10 GHz. No DC current, ideal resistors, and the 12 V side an AC ground.
Link: A cascade of each board’s trace, as a line with the impedance and delay from Antmicro’s own field-solver results, the 100 nF capacitor, the ESD parts’ capacitance, a stand-in for each FAKRA connector, and 15 m of coax. The coax’s loss has a part that grows with the square root of frequency and a part that grows with frequency, fitted to 0.9 dB/m at 3 GHz new and 1.1 dB/m aged. Limits from Equations 3 and 4 of the GMSL2 channel spec.
Camera’s pin: Random 6 Gb/s bits launched at 400 mV with 50 ps edges, and 187.5 Mb/s control bits launched at 250 mV with 2.3 ns edges, the data sheets’ typical figures. The camera hears its own signal, a −22 dB reflection near its connector and a −30 dB one from the far end. The canceler is a textbook LMS design spanning 48 video bits at eight phases, learning over 2,400 control bits, 12.8 µs. The receiver’s low-pass is second order at 150 MHz.
Opening: 20 ns near the end of that run, once the canceler has learned.
Sums: ASA Motion Link’s line rate and buffer above are worked from Kang and Gorshe’s burst timing. The gaps’ total is what the cycle leaves after the two bursts.

Simulated only. Nothing on this page has been measured on a board yet.

Limits

ADI publishes nothing on its canceler. The one here is a textbook design, so Fig. 27f shows what canceling can do, not what the MAX96717 does.
Random bits stand in for GMSL2’s 9b10b code, which is DC-balanced and puts less of the video’s power near DC.
The 22 µH stand-in resonates at 27.3 MHz, against 21 MHz typical on the MSS6132T-223’s data sheet.
DC current lowers a coil’s inductance and moves its resonance. ADI’s rule of thumb lets saturation current cut it by no more than 10%.
The FAKRA connectors are stand-ins, not field-solved.
The video’s eye at 6 Gb/s isn’t shown. Its loss and echo are.

Files

sim/site.pyEvery curve and number on this page, from the three sims below
sim/poc.pyThe coils’ models, fitted to Coilcraft’s curve for the whole filter
sim/channel.pyThe whole link’s loss and echo, and ADI’s limits
sim/echo.pyThe camera’s pin and the echo canceler
post/site_boards.pyThe board renders in Fig. 27, makers’ marks painted out
runs/layout/Both boards’ copper, read from their KiCad files

Sources

Like this note?

Rev 1 · 8 Nov 2026

Rev 18 Nov 2026First published

Each fix gets a line here, with the reader's name if they want credit.

Did it answer your question?

Want more on this?

Report a problem

What should the next note be? Ask for one