A Guide to SDI Capture for Edge AI

SDI capture is the process of bringing uncompressed video from a Serial Digital Interface (SDI) source, such as a broadcast, inspection or defense camera on coaxial cable, into a computer through a capture card or frame grabber, so it can be recorded, streamed or analyzed. Because SDI carries uncompressed video with near-zero transport latency over long, locking cable runs, it suits real-time edge AI as well as broadcast.

An SDI capture card or module receives the SDI feed and turns it into a data stream the host can use over PCIe, M.2, Mini PCIe or USB. Once ingested, the video can be recorded, streamed or passed through GPU pipelines for edge AI inference, defect detection, situational awareness or autonomous navigation. In industrial edge computing, SDI capture is the bridge between rugged cameras in the field and the computers that turn pixels into decisions.

This guide compares the SDI standards and link configurations, sets SDI against HDMI, GMSL and IP video, covers the capture interfaces and how many channels a host can take, explains where latency comes from, and shows how SDI capture fits NVIDIA Jetson and x86 edge AI systems.

Over the past three decades, SDI has grown from standard-definition broadcast signals into ultra-high-resolution video transport. Each generation of SD-SDI, HD-SDI, 3G-SDI, 6G-SDI and 12G-SDI raised the bandwidth for higher frame rates, greater color depth and larger images, and each is backward compatible with the ones before it.

All of them use BNC connectors on 75-ohm coaxial cable, and every standard can carry embedded audio and ancillary data, such as timecode, alongside the picture. Cable reach depends heavily on the cable grade and the receiver’s equalisation, so treat the distances as planning figures and check the camera and card datasheets for your run. Fiber converters extend any of them to kilometers where needed.

As formats grew, SDI changed not only in bit rate but in how signals are delivered. Single-link SDI is the simplest path; more demanding formats can split the picture across several synchronized links.

Quad-link sources come in two layouts, square division and two-sample interleave, and the capture card has to support the one the camera sends. Multi-channel 12G-SDI cards can usually combine four inputs into one quad-link 8K60 capture. Our 4K and 8K capture cards list the formats each card accepts.

SDI is one of several ways to bring cameras into an edge AI computer. The right one depends on cable length, the camera you can buy, latency and whether the signal is compressed on the way.

SDI wins where the camera already speaks it, where runs are long and where a locking connector and no compression matter, which is why it remains the default on broadcast, surgical and many EO/IR defense cameras. Our answers on SDI vs HDMI, GMSL2 and GigE Vision cameras compare each link in more depth, and our explainer on the BNC connector covers the connector itself.

The edge is not one form factor, it is many. From rugged industrial PCs to compact NVIDIA Jetson systems, SDI capture adapts to the host through PCIe, Mini PCIe, M.2 and USB, so uncompressed video can flow into neural networks at the edge whether the computer is a rackmount server or a small fanless box.

High-performance PCIe cards carry FPGAs or ASICs that handle scaling, color space conversion and frame synchronization on the card, which keeps the host CPU and GPU free for inference. M.2 cards have made SDI practical on Jetson and other compact edge AI computers where size and thermal design are tight, and Mini PCIe cards keep SDI capture alive in industries where system replacement cycles span decades.

The SDI bit rate is not what the host sees. The capture card decodes the signal and delivers frames in a pixel format, and that is what has to fit through the bus and into memory. A quick calculation avoids most bottlenecks.

Worked example. A Jetson AGX Orin vehicle computer takes four 1080p60 3G-SDI cameras through an M.2 four-channel card. At about 1 GB/s in total the stream fits a PCIe x4 M.2 socket comfortably, and the AGX Orin’s 204.8 GB/s memory bandwidth leaves room for inference on every feed. Move the same cameras to 4K and the stream rises to about 3 GB/s, which calls for a PCIe x8 card and a host with an x8 slot. Our answer on 4K resolution covers the formats, and the figures here are planning estimates before blanking and driver overhead.

In AI workloads, milliseconds matter. SDI video is uncompressed, so SDI latency is set not by transport over the cable but by the capture and processing chain inside the system. Every element of the pipeline, from the capture card to the bus to the GPU memory path, decides how quickly pixels become decisions.

High-performance SDI capture devices with FPGAs or ASICs offload color space conversion, scaling and synchronization at wire speed, reducing the load on the host CPU. Paired with newer PCIe generations, they give the bandwidth for multi-channel 4K or 8K feeds without bottlenecks, and M.2 modules bring the same capability into compact fanless edge systems at the point of capture.

With GPUDirect RDMA, where the capture card’s driver supports it, frames can bypass a staging copy and land directly in GPU-accessible buffers where CUDA pipelines and TensorRT engines run. On a Jetson, the CPU and GPU already share one memory, so the gain comes from avoiding extra copies rather than crossing a bus. Either way, the aim is a sub-frame budget from capture to decision. Our answer on RDMA for edge AI explains the mechanism.

Most SDI capture cards present a standard Video4Linux (V4L2) device on Linux and a DirectShow device on Windows, with a vendor SDK for multi-card, ancillary data and audio control. That makes them straightforward to bring into GStreamer, OpenCV, DeepStream or a TensorRT application.

The driver is the step to check first on Jetson. Yuan, for example, lists Linux Arm support and names Jetson TX2, Nano and Xavier modules for many of its M.2 and Mini PCIe SDI cards; for an Orin or Thor build, confirm a driver exists for the JetPack release you will run before you fix the design. Where SDI is central, rugged Jetson computers with SDI inputs built in avoid the question. For bigger channel counts, x86 edge AI servers with PCIe capture cards and a discrete GPU take more streams per system.

Our AI medical cart project shows the medical case with endoscopy video capture, and our answer on EO/IR cameras covers the defense sensors that often output SDI.

We configure each capture card in its host, install the driver for the operating system or JetPack release, and test every channel before the system ships.