System designExplainer

CSI virtual channels for multiple sensors on one port

How many CSI virtual channels one port can carry, the bandwidth multiple 1 MP RAW10 sensors need on two lanes, and the checks that decide whether they fit.

LumaSpec Vision11 min read Share on LinkedIn

On this page 9 sections
  1. 01How many sensors can share one CSI port?
  2. 02How much bandwidth does a 1 MP sensor at 15 fps need?
  3. 03Do four of those sensors fit on a two-lane port?
  4. 04What limits the sensor count in practice: lanes or drivers?
  5. 05How do you configure four sensors on one port?
  6. 06How do you verify all channels stream together?
  7. 07When should you move a sensor off CSI to USB3 or GigE?
  8. 08What should you check before you build the harness?
  9. 09Questions

Six small sensors, two CSI ports, and one port already promised to a large sensor: the question is whether the leftover lanes can carry the rest as virtual channels. Two numbers decide it. The first is how many virtual channels your processor's receiver actually implements per port. The second is how much payload the streams add up to on the lanes they share. This article works both numbers for the common case, four 1 MP RAW10 sensors at 15 fps on a two-lane port, then shows where bandwidth stops being the limit and driver support takes over. You get the arithmetic to size any similar split, a bring-up order that finds channel limits early, and the exit path to USB3 or GigE board cameras when the processor says no.

How many sensors can share one CSI port?

On the specification side, a CSI-2 port carries four virtual channels, so four sensors is the working ceiling per port. Your processor may implement fewer or more, and the number that counts is the one in its camera receiver documentation.

A virtual channel is a label inside the protocol, not extra wiring. Every CSI-2 packet starts with a data identifier that includes a channel field, so one physical port can interleave packets from several transmitters while the receiver sorts them back into separate capture streams. The original specification allocates two bits to that field, which is where the count of four comes from. Later specification revisions widen the field, but most embedded receivers still implement the original four per port. Treat four as the working number, and treat anything higher as a claim your processor's camera documentation has to make in writing.

Watch for a second, independent ceiling: many embedded modules cap the total number of simultaneous cameras regardless of how many channels each port allows. If the module lists four channels per port and four cameras in total, six sensors cannot all attach at any bandwidth. Both numbers live in the same place, the vendor's camera subsystem documentation, and both need to be in hand before you draw the flex board.

How much bandwidth does a 1 MP sensor at 15 fps need?

A 1280 × 800 RAW10 stream at 15 fps carries about 154 Mbit/s of pixel data. Once 20 % of the port is budgeted for protocol overhead, the planning number is 192 Mbit/s of lane time per sensor.

payload (Mbit/s) = width (px) × height (px) × frame rate (fps) × bits per pixel / 1,000,000 planned lane load (Mbit/s) = payload / 0.80

Worked example, one sensor:

  • 1280 px × 800 px = 1,024,000 px per frame
  • 1,024,000 px × 15 fps = 15.36 Mpx/s
  • 15.36 Mpx/s × 10 bit/px = 153.6 Mbit/s of payload
  • 153.6 / 0.80 = 192 Mbit/s of planned lane load

The 20 % allowance is not arbitrary padding. Every long packet carries a header and a CRC check, frames are bracketed by short packets, and the link drops into low-power states between packets, all of which costs lane time the pixel math never shows. Streams with many small packets pay more; long continuous lines pay less. Twenty percent is a sound planning default, and the receiver's bandwidth and error counters will confirm the real share during bring-up.

Do four of those sensors fit on a two-lane port?

Yes on bandwidth: four planned loads of about 192 Mbit/s total roughly 770 Mbit/s. That is about a quarter of the 3 Gbit/s a two-lane port offers when each lane runs at the 1.5 Gbit/s D-PHY ceiling.

Link Line rate Payload budget at the 80 % rule Four-stream planned load
2-lane CSI-2 at 800 Mbit/s per lane 1.6 Gbit/s 1.28 Gbit/s 768 Mbit/s, 60 % of budget
2-lane CSI-2 at 1.5 Gbit/s per lane 3.0 Gbit/s 2.4 Gbit/s 768 Mbit/s, 32 % of budget
4-lane CSI-2 at 1.5 Gbit/s per lane 6.0 Gbit/s 4.8 Gbit/s 768 Mbit/s, 16 % of budget
4 × 1920 × 1080 RAW10 at 30 fps on 2 lanes 3.0 Gbit/s 2.4 Gbit/s 3.1 Gbit/s, does not fit

The per-lane figures are the MIPI D-PHY rates receiver chips commonly implement, with 1.5 Gbit/s the usual per-lane ceiling and 800 Mbit/s a conservative planning value. Run the rate your processor's documentation lists for the specific port. Even at 800 Mbit/s per lane the four streams sit just under half the line rate, which is a sound design margin: embedded data lines, error recovery and a stream added later all come out of the slack. The last row is the flip side, 4 × (622 Mbit/s payload / 0.80) = 3.1 Gbit/s. The same four channels carrying 1080p streams at 30 fps simply do not fit, so sensor size and frame rate, never channel count, set the bandwidth ceiling.

How a lane rate turns into a delivered frame rate is its own subject; the guide on MIPI camera lane bandwidth: what limits frame rate walks that conversion.

What limits the sensor count in practice: lanes or drivers?

At 1 MP and 15 fps the lanes are barely working; what stops designs is the receiver. The processor must implement as many virtual channels per port as you wire, and many modules also cap the total number of cameras regardless of channels.

Three ceilings stack up, each from a different document. Channels per port come from the receiver hardware; total cameras per module come from the processor's pipeline count; per-lane rate comes from the PHY. A spec-sheet claim of eight cameras over virtual channels usually means the paper arithmetic works out. What implements the demultiplexing is the delivered driver, one capture path per channel, and that is where port-sharing designs stall in practice. Put the questions to the processor vendor in exactly this form: how many virtual channels per port, how many cameras per module in total, how many lanes per port, at what per-lane rate. Numbers in the answers, before the harness is drawn.

How do you configure four sensors on one port?

Each sensor's driver needs its own virtual channel number, the receiver pipeline needs one capture path per channel, and the port's lanes are wired to all sensors through an aggregator that merges the streams before the processor sees them.

Two architectures get you there. Some receivers merge channels natively: the sensors' transmitters all drive the same lane pair, each stamps its own channel index, and the receiver sorts the packets into separate capture nodes. Because transmitters share one link, they must also share the lane rate and frame timing, which in practice means identical sensor settings or sensors that can be slaved to a common clock. The second architecture puts an aggregator in front of the port, an FPGA or a deserializer built for the job: it accepts per-input settings, arbitrates the lane time and emits one uniform stream, which relaxes the matching requirement at the cost of another part to configure.

The channel index itself is a setting in each sensor's driver or its media pipeline description. The name and location differ by driver, but every multi-channel stack has one, and distinct values per sensor are the first thing to set.

How do you verify all channels stream together?

Bring channels up one at a time and watch delivered frame rate and error counters. A stack that streams four sensors alone but drops frames with all four running is telling you the receiver, not the wiring, is at its limit.

  1. Stream channel 0 alone and check the delivered frame rate at its capture node against the sensor's rating. If it is already short with nothing else running, the delivered-rate method from stereo cameras that under-report fps finds where the loss lives.
  2. Bring up the second sensor on a distinct channel index and confirm both nodes deliver together. Intermittent failures that appear now usually trace to lane signal integrity; lane alignment errors look exactly like random capture failures on one node.
  3. Declare channels 2 and 3. If the extra capture nodes fail to open, the fault is in the pipeline description rather than the silicon; the two-camera overlay fault is the pattern to expect.
  4. Run all four channels for at least 30 minutes with the receiver's error counters visible. CRC errors or dropped frames under full load point at bandwidth or timing; a black stream at low fps points at the driver stack.
  5. Repeat the run while the high-resolution sensor streams on its own port. Shared internal bandwidth inside the processor can still bite when each port's own math fits; the sensor-and-host tuning pair is the checklist for that stage.

A harness that passes all five has cleared both ceilings in the only test that counts, all streams running at once.

When should you move a sensor off CSI to USB3 or GigE?

Move a sensor off the CSI ports when the receiver documents fewer channels than sensors or caps total cameras: a 1 MP RAW10 stream at 15 fps is 154 Mbit/s of payload, light enough that a single GigE or USB3 board-level camera carries it with room to spare.

That is the honest exit, and at these stream sizes it is an easy one. The shop's board-level cameras are exactly this class of part: GigE and USB3 modules with the sensor on a board you build into your own housing.

Board-level camerasGigE board camera with split sensor, 1.0 MP, Mono (LS-BCM-0059)LS-BCM-0059
Sensor format
1/4"
Lens mount
M12
Interface
GigE
Resolution
Up to 3 MP
$209.00View partBuild a system with it

A camera like the LS-BCM-0059, a 1.0 MP mono board camera on GigE, moves its stream off the CSI port entirely and onto a link no other stream competes for.

Board-level cameras3.1 MP USB3 board-level camera, IMX252, 1/1.8" global shutter, 120 fps (LS-BCM-0003)LS-BCM-0003
Sensor format
1/1.8"
Lens mount
C-mount
Interface
USB3
Resolution
5 MP
$690.00View partBuild a system with it

The LS-BCM-0003 streams 3.1 MP at 120 fps over USB3, which covers the high-rate sensor case as well. Where the build is tight, both modules below stream 1.3 MP over USB3 or GigE in compact form factors.

Board-level camerasUSB3.0 board camera / module, 1.3 MP, Color (LS-BCM-0047)LS-BCM-0047
Sensor format
1/2.7"
Lens mount
C-mount
$179.00View part
Board-level camerasMiniature industrial camera, 1.3 MP, GigE, Colour/Mono (LS-BCM-0061)LS-BCM-0061
Sensor format
1/2.7"
Lens mount
M12
$209.00View part

Budget the host side as you move streams off CSI. A single GigE link carries the 154 Mbit/s stream with most of its payload budget free. The per-camera arithmetic for USB3 links is laid out in the guide on reducing USB3 camera bandwidth, and if the USB3 cameras plug through a hub, budget the hub first. When several GigE cameras land on a single host port, budget NIC bandwidth and CPU together. Run each new link on camera data cables rated for the run length and any flex duty.

What should you check before you build the harness?

Before any PCB or flex is ordered, have these answers in writing: they are the difference between a harness that works on day one and a redesign two months in.

From the processor documentation

  • Virtual channels implemented per CSI port: the hard ceiling on sensors per port; four is the common figure.
  • Total simultaneous cameras per module: caps the whole design even when per-port channels are free.
  • Lanes per port and the documented per-lane rate: feeds the bandwidth math.

From your own stream math

  • Payload per stream from the formula, planned load at the 80 % rule, summed per port, kept under half the line rate.
  • Pixel depth and frame rate actually set: RAW8 instead of RAW10 cuts the total by 20 %, and frame rate moves it linearly.

From the driver stack

  • One capture path per channel exposed by the receiver driver, confirmed in the driver's documentation rather than the marketing sheet.
  • A distinct channel index per sensor, and matching lane timing for transmitters that share lanes directly.
  • An aggregator, if used, documented for your channel count and lane rates.

From the bench plan

  • The five-step bring-up above, with delivered frame rate and error counters recorded at each step.

When the numbers clear, order sensors and draw the harness; when they do not, the same sensors on USB3 or GigE board-level modules are the lower-risk path. The camera finder filters by resolution, interface and shutter, and the system configurator walks the chain from sensor to cable. For a second opinion on a port-sharing plan, send the stream table through the contact page, or browse the systems guides for the other interface-budget articles.

Questions

Can I mix different resolutions and frame rates on one CSI port?

Every CSI-2 packet carries its own channel index and data type, so a 1080p stream and a 1 MP stream can interleave on the same lanes while the receiver sorts them into separate capture paths. Budget the sum of all planned loads for the port and check that each sensor's driver tolerates a shared port, since some sensor drivers assume they own the link.

How many virtual channels does the CSI-2 specification allow?

The base specification allocates four channels per port, and revision 2.0 raises the ceiling to 32 through the channel extension field. Most embedded receiver silicon and drivers still implement the original four, so plan around four unless the processor's camera documentation states a higher figure explicitly.

Do sensors sharing one port need hardware sync?

For independent inspections, no: each channel is its own stream. When the application compares images across sensors, as in stereo depth or multi-angle measurement, free-running clocks drift apart. Wire a shared trigger or strobe line so all sensors expose together and cross-sensor timestamps stay usable.

How does the math change for RAW8 instead of RAW10?

Replace 10 with 8 in the formula. The same 1280 × 800 sensor at 15 fps then carries 122.9 Mbit/s of payload, about 154 Mbit/s of planned lane load at the 80 % rule, roughly 20 % below the RAW10 figure this article uses.

Does sharing one port add latency to the streams?

Packets from all channels interleave on the lanes, so a sensor sending a large frame can briefly delay another channel's packets. At these loads the effect is small. Measure delivered frame rate per channel with all streams running to confirm no channel loses more than its share.