How does EtherCAT reach such short cycle times over ordinary Ethernet cables and ports?

The difference is that the slaves do not receive and answer frames individually. The master sends one Ethernet frame (EtherType 0x88A4) that travels through all slaves in turn. Each slave has two or more ports and a dedicated EtherCAT slave controller chip, which reads the data addressed to it and inserts its own input data while the frame is passing through, in hardware, with a delay in the region of a microsecond per device. The last slave sends the frame back, and it returns to the master through the same chain.

So one frame, sent once per cycle, carries the outputs for every slave and comes back containing every input.

Some numbers show why this is fast. A maximum-size Ethernet frame with 1500 bytes of payload occupies the wire for about 123 µs at 100 Mbit/s, counting preamble and inter-frame gap, and it can carry the process data of a whole machine. Compare Modbus TCP, where every device needs its own request and response through a software TCP/IP stack on both sides.

The addressing helps too: the master maps the process data of all slaves into one logical image, and a single read/write command in the frame serves every slave, each picking out its own few bytes.

Hardware: the master needs nothing special, just a standard Ethernet port and a master stack that sends raw Ethernet frames with good timing, normally under a real-time operating system or a real-time extension. Open-source stacks such as SOEM and the IgH EtherCAT master exist for learning.

The slaves are what need dedicated hardware. Every slave contains an EtherCAT slave controller (an ASIC, an FPGA core, or a microcontroller with the controller built in), because the on-the-fly processing cannot be done in software. The physical layer is ordinary 100BASE-TX: standard Cat 5e cable, up to 100 m between two devices.

Do not put slaves behind an ordinary switch. A standard switch stores and forwards frames by MAC address, while an EtherCAT segment is a chain in which every device modifies the passing frame; the usual arrangement is a direct cable from the master's port to the first slave. Branches are made with EtherCAT junction devices, which are themselves slaves with extra ports.

Topology is flexible because of how the ports work: a slave that detects no link on a port closes that port internally and sends the frame back. A line, a tree with branches and drop lines are all the same logical ring to the master. Addresses are normally assigned by position at start-up, so devices need no address switches, and up to 65,535 devices fit in one segment.

For a first project, the steps are: get the ESI file (the XML device description) for each slave from its manufacturer, let the master tool scan the bus and compare what it finds with the configuration, then bring the slaves from INIT through PRE-OP and SAFE-OP to OP. In OP the cyclic frame exchanges process data.

A useful diagnostic from the start is the working counter returned with each datagram. Every slave that successfully handled the command increments it, so the master knows in the same cycle whether all expected slaves took part.