IP implements packet-switched networking. It has a concept of hosts, which are machines. The IP protocol specifies how datagrams (packets) are sent between these hosts.

A packet is a chunk of binary data that has a source host and a destination host. An IP network will then simply transmit the packet from the source to the destination. One important aspect of IP is that packets are delivered using the best effort. A packet may be lost along the way and never reach the destination. Or it may get duplicated and arrive multiple times at the destination.

Each host in an IP network has an address -- the so-called IP address. Each packet contains the source and destination hosts' addresses. The IP is responsible for routing datagrams -- as the IP packet travels through the network, each node (host) that it travels through looks at the destination address in the packet to figure out in which direction the packet should be forwarded.

Today, most packages are still IPv4 (Internet Protocol version 4), where each IPv4 address is 32 bits long. They're most often written in dotted-decimal notation, like so: 198.51.100.42

The newer IPv6 standard is slowly gaining traction. It has a larger address space: its addresses are 128 bits long. This allows for easier routing as the packets travel through the network. And since there are more available addresses, tricks such as network address translation are no longer necessary. IPv6 addresses are represented in the hexadecimal system and divided into eight groups separated by colons, e.g., 2001:0db8:85a3:0042:1000:8a2e:0370:7334

An IP packet consists of a header and a payload.

The header is 20 bytes long (without options, which are rarely used).

The IPv6 header has a fixed length of 40 bytes. It's a lot simpler than IPv4 - a few lessons were learned in the years that have passed since IPv4.

In IPv4, packets (datagrams) can get fragmented. The underlying transport layer will have an upper limit to the length of a packet it can support. In IPv4, a router may fragment a packet if it gets routed onto an underlying data link for which the packet would otherwise be too big. These packets will then get reassembled at the destination host. The sender can decide to disallow routers to fragment packets, in which case they'll send a Packet Too Big ICMP back to the sender.

In IPv6, a router will always drop the packet and send back a Packet Too Big ICMP6 message to the sender. The endpoints use this to make a path MTU discovery to figure out what the optimal so-called maximum transfer unit (MTU) along the path between the two hosts is. Only when the upper layer has a minimum payload size that is too big for this MTU will IPv6 use fragmentation. With TCP over IPv6, this is not the case.