What Is MIPI CSI-2?

MIPI CSI-2, the Camera Serial Interface 2, is the specification that defines how an image sensor sends pictures to a processor. It is published by the MIPI Alliance, the industry body formed to standardise interfaces inside mobile devices (MIPI stands for Mobile Industry Processor Interface). The first version appeared in 2005, and it is now the usual camera input on smartphone processors and on embedded processors such as NVIDIA Jetson, NXP i.MX, Rockchip and Raspberry Pi.

A CSI-2 link is one-way and packet based. The sensor splits each frame into lines and each line into packets with a header and error checking, then sends them over one to four high-speed data lanes alongside a clock. A separate low-speed control channel, compatible with I²C, lets the processor set exposure, gain and frame rate. Virtual channels let one link carry several streams, such as two exposures for high dynamic range or the outputs of several sensors behind a serialiser.

Because the link is so direct, there is very little in the way: the raw sensor data goes straight into the processor’s image signal processor (ISP) for demosaicing, noise reduction and colour correction, with low latency and low power.

CSI-2 sets the protocol. The electrical signalling underneath comes from separate MIPI physical layer specifications, and the choice sets the speed and the distance.

The main limit of MIPI CSI-2 is distance. D-PHY was designed for the few centimetres inside a phone, so in embedded systems the camera normally sits on a flat flexible cable of around 30 cm, and some interface boards stretch that to about 50 cm with careful design. Connectors are fine-pitch FPC or board-to-board types; the Raspberry Pi’s 15-pin and 22-pin ribbon connectors are the best-known examples.

For longer runs, the CSI-2 stream is serialised onto coax or fibre and converted back near the processor. GMSL from Analog Devices and FPD-Link from Texas Instruments both work this way, and the processor still sees an ordinary CSI-2 input. Our GMSL cameras use that approach to reach 15 m on a vehicle or robot, and our guide to GMSL technology for industrial edge AI explains the serialiser chain in detail.

CSI-2 is one of four camera interfaces commonly found on industrial and edge AI computers. The right one depends mostly on the distance between camera and computer and how many cameras the system needs.

CSI-2 has the lowest latency, power and cost per camera, but it ties the camera to the processor’s own interface and drivers. GigE and USB cameras are interchangeable between computers, while a CSI-2 sensor needs a driver written for that processor. Whatever the interface, sensor choice matters as much; our article on global shutter vs rolling shutter covers the effect on moving objects.

NVIDIA Jetson modules have CSI-2 inputs built in, so they are the most common home for MIPI cameras in edge AI. Depending on the carrier board, Jetson AGX Orin and Jetson Thor expose up to 16 lanes, and Orin NX and Orin Nano up to 8, enough for several cameras with virtual channels. Our Jetson Orin carrier boards list the lane count and camera connectors for each design, and compact computers such as the JCO-1000-ORN series include internal MIPI CSI-2 connectors.

A MIPI camera only works if the Jetson software knows the sensor. Each one needs a kernel driver and device tree entries in the board support package, and raw sensors are tuned for the Jetson ISP. Camera makers supply these for specific modules and JetPack releases, so match the camera, carrier board and JetPack version before you build.

We supply our Jetson computers with cameras, drivers and the JetPack image matched and fitted as standard, and check every camera input before the system ships.