GMSL vs USB for depth cameras: latency, cable length and noise

Put a depth camera more than a few metres from its host and the interface stops being a detail. USB 3 was built for desk distance, GMSL over FAKRA coax was built to cross a vehicle, and GigE or 10GigE was built for a switched plant network. This comparison gives the line rates, passive reach, link-latency behaviour, interference profile and power budget of all three, then works a 10 m, 90 fps example end to end so you can order the right link the first time. The decision rule is reach first, stream budget second; both come with numbers.

The decision, per cable run

Pick the interface by run length, then check the budget:

• Under 5 m: USB 3.0. One passive high-flex cable, no extra hardware, 4.5 W of port power. Retention is the only weak point, and only on a moving axis.
• 5 m to about 15 m: GMSL2 or GMSL3 over FAKRA coax. The serializer moves next to the sensor, so one latched 75 Ω coax carries a 6 Gbps (GMSL2) or 12 Gbps (GMSL3) pipe, usually with power over the same cable.
• Beyond 15 m, or several cameras on one host: GigE or 10GigE. 100 m on Cat6A, switched topology, a NIC or frame grabber at the receive end.

Reach picks the interface; the stream budget can veto it. A 6 Gbps GMSL2 pipe cannot carry a 7 Gbps stream set, and 1 Gbps Ethernet carries very few depth cameras at all — the worked example below puts numbers on that veto.

Interface numbers side by side

Two payload figures come straight from line coding, and they are the numbers to budget against:

GigE lands lower still: a UDP depth stream on a 1000BASE-T link clears about 0.94 Gbps once protocol overhead is paid. Most multi-stream depth cameras do not fit that, which is why plain GigE rarely appears in a depth-camera build.

Latency: what the link adds, and what it does not

At 30 fps one frame lasts 33.3 ms (1000 / 30); at 90 fps, 11.1 ms (1000 / 90). Exposure and readout eat most of that period no matter the interface, so the link's job is to add little on top — and, more importantly, to add the same amount every frame.

USB 3.0 schedules transfers in 125 µs microframes. A bulk depth stream usually clears scheduling in well under 1 ms, but the figure moves with host load, and every hub hop adds a few hundred microseconds more. GMSL's pipe is point-to-point with no arbitration and no retries: its delay is a fixed pipeline, constant frame to frame. At equal camera settings the end-to-end difference between the links is at most a few milliseconds; what separates them is that GMSL's delay never varies. A picking or servo loop profits from that determinism more than from a marginally lower average.

Cable reach and the physical run

USB 3.0 passive cables are specified to about 3 m; 5 m works on a clean, well-shielded run and is the practical ceiling. Past that you are buying active optical cables or powered repeaters, and each is one more powered conversion stage that can drop a stream. That failure mode is behind most USB cable length limits on 3D cameras and USB-C timeouts that appear only under load.

GMSL SerDes kits ship 5 m and 10 m FAKRA coax leads, and GMSL2 at 6 Gbps is quoted to about 15 m on quality 75 Ω coax. The connector matters as much as the length: a FAKRA snap latch is built for automotive vibration, while USB retention is friction and a light click. 10GBASE-T runs 100 m on Cat6A, which makes Ethernet the answer for anything across a cell or down a line; the general interface trade is laid out in the comparison from USB to CoaXPress.

Noise: radiated and received

A 5 Gbps USB 3.0 line rate concentrates spectral energy in the low-GHz range, and any break in the shield — a low-quality moulding, an unshielded pigtail inside the cabinet — radiates it into the 2.4 GHz band where Wi-Fi and Bluetooth live. The classic symptom is radio dropout whenever the camera cable passes an antenna. Shielded high-flex cable and short internal runs fix it.

GMSL moves the fast edge to the sensor. The serializer sits centimetres from the imager, so the only high-speed digital section is a stub inside the camera head; the long run is coax, which is built to carry RF without radiating it, inside latched, shielded connectors. Less noise leaves the cable and less couples into it, which is why intermittent drops on a long serial link get hunted at the cable first — the method is in isolating a GMSL camera that stops streaming.

Power over the link

A USB 3.0 port supplies 5 V at up to 0.9 A — 4.5 W. Cameras with active projectors routinely draw more, and the symptoms are black frames and disconnects; the diagnosis is in a black depth camera image traced to USB bandwidth and power, and the same trade-off applies to any external 5 V supply you substitute for the port. PoE covers the Ethernet case at the 15.4 W class per port.

GMSL designs commonly use power over coax: one FAKRA cable then carries data and power together, and the budget moves to the host-side supply behind the deserializer. Both ends must support PoC for that to hold. Size the host input from the camera's measured draw — the working method is in the GMSL camera power budget check.

Worked example: a 10 m run to a 90 fps depth camera

Inputs. Streams: depth 1280 × 720 at 90 fps, 16-bit; IR 1280 × 720 at 90 fps, 8-bit; RGB 1280 × 720 at 30 fps, 8-bit — a typical mid-range depth camera set, so substitute your own resolution and rates. Run: 10 m from a camera on a mast to a fixed host. Line rates are the interface standards: USB 3.0 5 Gbps, GMSL2 6 Gbps, 10GigE 10 Gbps.

Rule. One stream costs width × height × bytes per pixel × 8 × frame rate in bits per second. Budget the sum of the streams plus 10 % for protocol and blanking overhead.

Arithmetic.

• Depth: 1280 × 720 × 2 × 8 × 90 = 1,327,104,000 ≈ 1.33 Gbps
• IR: 1280 × 720 × 1 × 8 × 90 = 663,552,000 ≈ 0.66 Gbps
• RGB: 1280 × 720 × 1 × 8 × 30 = 221,184,000 ≈ 0.22 Gbps
• Sum 2.21 Gbps, × 1.1 → 2.43 Gbps budget

Decision.

• USB 3.0: 2.43 of 4.0 Gbps payload — 61 % used, 39 % headroom. Bandwidth fits; reach does not, because 10 m is twice the practical passive limit and forces an active segment.
• GMSL2: 2.43 of 6 Gbps — 41 % used, 59 % headroom, and 10 m is a standard kit length. One latched coax does the job.
• GigE: 2.43 of 0.94 Gbps fails outright. 10GigE: 2.43 of 9.7 Gbps fits with 100 m of reach, at the cost of a 10G NIC and more host memory traffic.

Threshold: below 5 m take USB 3.0; from 5 m to about 15 m at under 6 Gbps of streams take GMSL2, with GMSL3 lifting the ceiling to 12 Gbps in the same reach class; beyond that, 10GigE.

Host side: the receive end

Whatever the camera speaks, the host needs a port that speaks it back. USB wants a root port that delivers its 4.5 W; 10GigE wants a PCIe NIC; GMSL wants capture inputs that many hosts simply do not have — embedded compute boards carry them, and bring-up is a project of its own, as in configuring GMSL and FPD-Link III camera access on an embedded host. If the host has no such input and none can be added, GMSL is out however well it fits the reach — plan Ethernet instead, and expect a mis-configured serial link to announce itself as VI timeouts and green screens.

As cards, three receive ends cover the common cases. The MV-GU2204 adds four USB3 ports to a bench whose onboard roots are taken. The MV-GT2104 is a quad-port copper 10GbE NIC for the long-run cameras. The MV-GE2002P streams and powers two PoE cameras at GigE.

All 3 share: Humidity 10% RH to 95% RH (no condensation). Values as printed in each product's data sheet; open a part for its drawing and full specifications.

Checks before you order

1. Measure the routed path, not the machine drawing. Add 20 % for dressing, service loops and the connector approach angle.
2. List every stream you will enable at production rates — depth, both IR feeds if stereo, RGB — and apply the × 1.1 overhead from the worked example. Preview-only sums are why long USB extension cables fail only when you record the depth stream.
3. Confirm the receive end before the camera: a powered USB port, a GMSL capture input, or a free PCIe slot and NIC for 10GigE.
4. Check power against draw. Over 4.5 W on USB means an external supply; beyond the PoC budget on GMSL means a separate power lead.
5. Match the cable class to the motion. Anything on an axis gets drag-chain rating; anything vibrating gets a latched connector. For a short moving USB run, a 3 m high-flex USB3.0 cable is the part class to order:

When the usual answer is wrong

• A 2 m run does not need GMSL. A SerDes pair adds two parts and a bring-up project to a run one passive cable already covers.
• A short run on a vibrating axis still wants GMSL — there you are buying the FAKRA latch, not the reach.
• Several cameras per host favours Ethernet. GMSL pipes are strictly point-to-point, so each camera consumes its own capture input; a switched 10GigE network feeds many cameras into one NIC, and the multi-camera USB contention behind the interface-claim error when two depth cameras share the bus disappears with it.

Sort the stream list, measure the run, and the interface picks itself. Depth cameras sit in 3D sensors with their camera finder; the receive end lives in cables and frame grabbers, and the build-a-system configurator walks the chain end to end. More link-level troubleshooting is collected on the systems topic page. Send your stream list and measured run through the contact page and we will check this arithmetic against your build.