Embedded camera interface comparison: from CSI-2 to GMSL2

Embedded vision gives you four realistic camera links, and they separate on two numbers: how far the cable runs and how much image data it moves per second. MIPI CSI-2 is the fastest but the shortest. GMSL2 and FPD-Link III carry multi-gigabit payloads across 10-15 m of serialized cable. USB3 Vision balances the two at 5 m. This comparison gives you the planning limits for each link, one worked sizing example with the arithmetic shown, and a decision for every run length, so you can commit before you order cameras, boards and cable.

Decide from two numbers: reach and payload

Size every link the same way. Compute the image payload at your worst-case pixel format. Add a 10% margin, then pick the shortest link that clears the total. The planning limits to design around:

• Camera within about 0.5 m of the host processor: MIPI CSI-2, up to about 10 Gb/s.
• Runs to 5 m with a payload under about 3.6 Gb/s: USB3 Vision.
• Runs of 5-10 m: FPD-Link III, about 4.26 Gb/s of forward-channel payload.
• Runs to 15 m: GMSL2, about 6 Gb/s.
• Beyond 15 m, or several cameras sharing switches: GigE-class Ethernet, at 1, 5 or 10 Gigabit.

Chipset choice and pixel format move each figure. The sections below show where each limit bends and what it costs to ignore.

The four links at a glance

MIPI CSI-2 is the camera-to-processor standard inside embedded systems. Four D-PHY lanes at 2.5 Gb/s each add up to about 10 Gb/s of payload. Power draw and electromagnetic interference are low, and the electrical spec is written for board-level runs rather than factory harnesses. If the camera module itself is still an open choice, the board-level camera selection guide covers the sensor trade-offs that come first.

USB3 Vision wraps USB 3.0 in the GenICam control model, so the camera behaves like its GigE sibling: discover, set exposure, stream, with power on the same cable. The 5 Gb/s line rate leaves about 3.6 Gb/s for image data once overhead is paid, and the passive cable stops at 5 m. For the host side of that decision, choosing an ARM board for USB3 camera acquisition walks through the controller questions.

FPD-Link III serializes the camera stream, which at the sensor is usually CSI-2, onto a single coax or twisted pair. It was built for automotive temperature ranges and electrical transients, which is why it survives factory cells. Its forward channel carries about 4.26 Gb/s to about 10 m, with electromagnetic emissions kept tight by design. The integration trap is bring-up without a board guide; FPD-Link III bring-up when no board guide exists shows how to isolate it.

GMSL2 is an asymmetric full-duplex serial link: high-rate video downstream, with control data and power moving in both directions over the same cable, to about 15 m at about 6 Gb/s. When such a link comes up dead, power and cabling are the usual suspects, as when a GMSL camera is not detected details.

Reach: what limits each run

Three things bend the planning reaches. Cable quality bends them up or down: good shielded twisted pair or quality coax holds a serialized link to its full length, while a marginal cable shows errors first. Flexing pulls them down: any harness that rides a drag chain needs high-flex construction regardless of interface. Bridges change the picture entirely. A CSI camera Ethernet extension kit carries a CSI-2 stream across standard Ethernet hardware, and the CSI extension cable checks list what to confirm before trusting one. USB 3.0 has no equivalent that keeps the protocol alive, which is why USB cable length limits matter for any long run, not just 3D rigs.

Payload: the sizing formula

Payload is arithmetic, not a datasheet lookup:

The 1.1 factor is the margin that absorbs control packets, trigger jitter and frame bursts. Skip it and you get the classic bench-versus-cell failure: clean frames at the desk, dropped frames beside a running machine. Do the sum with the unpacked bit depth (10 bit/px, not a packed variant) and the result stays on the safe side. The same budgeting habit, applied at switch level, is what machine vision Ethernet bandwidth planning formalizes for multi-camera hosts.

For how these limits line up at 4K and 60 fps, the 4K machine vision at 60 fps: bandwidth and interface limits analysis goes link by link.

Worked example: sizing the link for a 5 MP camera at 60 fps

Inputs: a mono 2448 × 2048 sensor, the standard 5 MP global-shutter class; 8-bit output at 60 fps full frame, the common starting format; run length undecided, because that is what this calculation decides. The rule from the last section applies: budget = 1.1 × pixels × depth × frame rate.

• Pixels per frame: 2448 × 2048 = 5,038,848 px
• Bits per frame: 5,038,848 px × 8 bit/px = 40,310,784 bit ≈ 40.3 Mbit
• Payload: 40,310,784 bit × 60 fps = 2,418,647,040 bit/s ≈ 2.42 Gb/s
• Budget: 1.1 × 2.42 Gb/s ≈ 2.66 Gb/s

Result: the link must carry about 2.66 Gb/s. Headroom against each planning limit: (3.6 - 2.66) / 3.6 ≈ 26% on USB3 Vision, (4.26 - 2.66) / 4.26 ≈ 38% on FPD-Link III, (6 - 2.66) / 6 ≈ 56% on GMSL2, (10 - 2.66) / 10 ≈ 73% on CSI-2. A 1 Gigabit Ethernet port fails outright; the workload is 2.7 times the link.

Decision: order USB3 Vision for any run up to 5 m, FPD-Link III for 5-10 m, GMSL2 for 10-15 m, and CSI-2 when the camera sits within 0.5 m of the processor. Moving the same sensor to 10-bit output raises the payload to 5,038,848 × 10 × 60 = 3,023,308,800 bit/s ≈ 3.02 Gb/s, or 3.33 Gb/s with margin. USB3 headroom then drops to about 8%, so FPD-Link III or GMSL2 becomes the safer pick at any real distance.

Side by side

Each figure below is the value to budget with, at the lane counts and chipsets that dominate current designs:

Which interface, case by case

Your own instrument, camera on or beside the carrier board: CSI-2, up to about 10 Gb/s at the lowest power of the four. Hosts without native MIPI input can still take the stream; capturing HDMI output through a CSI-2 interface on an embedded host walks one bridge path.

One camera, 1-5 m, standard PC, fastest integration: USB3 Vision. Two USB3 board-level cameras that suit this build style:

The LS-BCM-0003 pairs a 3.1 MP IMX252 global-shutter sensor with USB3 at 120 fps. The MV-CB060-10UM-B runs a 6 MP IMX178 sensor on a bare board you mount behind your own front panel.

Runs of 5-10 m beside drives and servos: FPD-Link III, for the transient tolerance and the controlled emissions.

Runs of 10-15 m, or video, control and power on one cable: GMSL2, the longest planning reach of the four.

Longer than 15 m, several cameras, switches already in the wall: GigE-class Ethernet. At 1 Gigabit the port is the constraint for current sensors; the 5GBASE-T camera interface guide and the GigE Vision selection guide cover the upgrade path. The GigE counterpart to the board cameras above:

The MV-CB016-10GM-M-W is a 1.6 MP GigE board camera with an M12 lens mount and a global shutter. All the board cameras named above, side by side:

Values as printed in each product's data sheet; open a part for its drawing and full specifications.

Power, noise and the cable itself

Power divides the camps. GMSL2 and FPD-Link III serializer pairs commonly carry power along the same coax or twisted pair as the video, which removes a second cable from the harness; total the camera's draw against what the chipset pair supplies before you rely on it. USB 3.0 bus power covers small board cameras but not a camera driving its own lighting, so plan a local supply wherever LEDs hang off the camera.

Electrical noise follows the same split. The serialized links were engineered for harnesses that live next to motors and drives; routing them along a cabinet's power trunking is normal duty. CSI-2 keeps emissions low by running short and at modest pin rates. USB3 wants a shielded cable and some separation from motor cables, and nothing more exotic than that.

Where the harness moves or rides a drag chain, high-flex construction separates a cable that lasts years from one that fatigues in months:

Both are 3 m high-flex cables, one built for GigE cameras and one for USB3.0.

Checks before you order

1. Compute the payload at the format you will actually run: 10- or 12-bit output multiplies the 8-bit figure by 1.25 or 1.5, and colour doubles it.
2. Measure the real cable route, including vertical drops and the service loop, not the straight line on the layout drawing.
3. Confirm the host input and its driver support: MIPI lanes on the carrier, a proven USB 3.0 controller, or a deserializer matched to your serializer.
4. Settle the power budget: bus power, link power from a serializer pair, or a local supply at the camera.
5. Choose the flex rating from the machine, not the wish; any harness that moves gets drag-chain-rated cable.
6. Order one spare cable of each type you deploy, and keep the bring-up guides for FPD-Link III and GMSL at hand for the first power-on.

When the usual answer is wrong

"USB everywhere" fails two ways: silently past the passive cable limit, and noisily when several cameras exhaust a shared host controller before the spreadsheet says they should. USB3 Vision transfer errors shows what that second failure looks like in the logs.

"CSI-2 because it is fastest" fails at distance: past board-level reach, a camera more than an arm's length away needs a serializer anyway, and then the real choice is FPD-Link III or GMSL2.

"GigE because the plant already has switches" holds only while the sensor is small. The worked example's 2.66 Gb/s budget is 2.7 times what a 1 Gigabit port carries, and no camera setting recovers the difference; the fix is 5GBASE-T, 10GigE, or a shorter link.

If you have a sensor, a frame rate and a run length on the desk, the camera finder and the build-a-system configurator filter by interface and resolution so you can shortlist parts today, or contact the applications team with the numbers and we will size the link with you. The systems topic page gathers the rest of the interface and bandwidth guides.