USB communication takes the form of packets. Initially, all packets are sent from the host, via the root hub and possibly more hubs, to devices. Some of those packets direct a device to send some packets in reply.

After the sync field described above, all packets are made of 8-bit bytes, transmitted least-significant bit first. The first byte is a packet identifier (PID) byte. The PID is actually 4 bits; the byte consists of the 4-bit PID followed by its bitwise complement. This redundancy helps detect errors. (Note also that a PID byte contains at most four consecutive 1 bits, and thus will never need bit-stuffing, even when combined with the final 1 bit in the sync byte. However, the OUT PID byte ends with three consecutive 1 bits, so if the following USB device address begins with three 1 bits, bit-stuffing will be required.)

Packets come in three basic types, each with a different format and CRC (cyclic redundancy check):

Handshake packets consist of nothing but a PID byte, and are generally sent in response to data packets. The three basic types are ACK, indicating that data was successfully received, NAK, indicating that the data cannot be received at this time and should be retried, and STALL, indicating that the device has an error and will never be able to successfully transfer data until some corrective action (such as device initialization) is performed.

USB 2.0 added two additional handshake packets, NYET which indicates that a split transaction is not yet complete, and an ERR handshake to indicate that a split transaction failed.

The only handshake packet the USB host may generate is ACK; if it is not ready to receive data, it should not instruct a device to send any.

Token packets consist of a PID byte followed by 11 bits of address and a 5-bit CRC. Tokens are only sent by the host, never a device.--

IN and OUT tokens contain a 7-bit device number and 4-bit function number (for multifunction devices) and command the device to transmit DATAx packets, or receive the following DATAx packets, respectively.

An IN token expects a response from a device. The response may be a NAK or STALL response, or a DATAx frame. In the latter case, the host issues an ACK handshake if appropriate.

An OUT token is followed immediately by a DATAx frame. The device responds with ACK, NAK, or STALL, as appropriate.

SETUP operates much like an OUT token, but is used for initial device setup.

Every millisecond (12000 full-speed bit times), the USB host transmits a special SOF (start of frame) token, containing an 11-bit incrementing frame number in place of a device address. This is used to synchronize isochronous data flows. High-speed USB 2.0 devices receive 7 additional duplicate SOF tokens per frame, each introducing a 125 µs "microframe".

USB 2.0 added a PING token, which asks a device if it is ready to receive an OUT/DATA packet pair. The device responds with ACK, NAK, or STALL, as appropriate. This avoids the need to send the DATA packet if the device knows that it will just respond with NAK.

USB 2.0 also added a larger SPLIT token with a 7-bit hub number, 12 bits of control flags, and a 5-bit CRC. This is used to perform split transactions. Rather than tie up the high-speed USB bus sending data to a slower USB device, the nearest high-speed capable hub receives a SPLIT token followed by one or two USB packets at high speed, performs the data transfer at full or low speed, and provides the response at high speed when prompted by a second SPLIT token. The details are complex; see the USB specification.

Data packets
There are two basic data packets, DATA0 and DATA1. Both consist of a DATAx PID field, 0–1023 bytes of data payload (up to 1024 in high speed, at most 8 at low speed), and a 16-bit CRC. They must always be preceded by an address token, and are usually followed by a handshake token from the receiver back to the transmitter. The two packet types provide the 1-bit sequence number required by Stop-and-wait ARQ. If a USB host does not receive a response (such as an ACK) for data it has transmitted, it does not know if the data was received or not; the data might have been lost in transit, or it might have been received but the handshake response was lost.

To solve this problem, the device keeps track of the type of DATAx packet it last accepted. If it receives another DATAx packet of the same type, it is acknowledged but ignored as a duplicate. Only a DATAx packet of the opposite type is actually received.

When a device is reset with a SETUP packet, it expects a DATA0 packet next.

USB 2.0 added DATA2 and MDATA packet types as well. They are used only by high-speed devices doing high-bandwidth isochronous transfers which need to transfer more than 1024 bytes per 125 µs "microframe" (8192 kB/s).

PRE "packet"
Low-speed devices are supported with a special PID value, PRE. This marks the beginning of a low-speed packet, and is used by hubs which normally do not send full-speed packets to low-speed devices. Since all PID bytes include four 0 bits, they leave the bus in the full-speed K state, which is the same as the low-speed J state. It is followed by a brief pause during which hubs enable their low-speed outputs, already idling in the J state, then a low-speed packet follows, beginning with a sync sequence and PID byte, and ending with a brief period of SE0. Full-speed devices other than hubs can simply ignore the PRE packet and its low-speed contents, until the final SE0 indicates that a new packet follows.