1. The Frame Hierarchy

An $\\\\text{Ethernet}$ frame is the Protocol Data Unit ($\\\\text{PDU}$) of the Data Link Layer ($\\\\text{Layer 2}$). It exists purely to ensure your data gets from one physical port to another across a single building or campus.

2. Visual Breakdown of an Ethernet II Frame

3. Component Analysis

• Preamble: A pattern of alternating 1s and 0s that "wakes up" the receiving hardware and synchronizes the internal clock speed.
• SFD (Start Frame Delimiter): The sequence$10101011$ which signals that the very next bit is the start of the address.
• $\\\\text{EtherType}$: This tells the $\\\\text{OS}$ whether it should hand the data to the $\\\\text{IPv4}$ stack, $\\\\text{IPv6}$ stack, or $\\\\text{ARP}$ handler.
• $\\\\text{FCS}$ (Frame Check Sequence): The error-checking heart of the frame. If a single bit is flipped by electrical interference, the $\\\\text{FCS}$ will catch it.

4. The 64-Byte Minimum Rule

Why can't a frame be 20 bytes? In the early days of shared-medium Ethernet (Hubs), a sender had to keep talking long enough for its signal to reach the furthest end of the cable and back.

5. Baby Giant and Tagged Frames

When we add a $\\\\text{VLAN}$ tag (802.1Q), we add $4$ extra bytes to the header. This increases the frame size beyond the standard limit. Modern switches are designed to handle these "Baby Giant" frames ($1504\\,\\\\text{bytes}$ or $1518\\,\\\\text{bytes}$ total) without error.

6. PAM-4 and 802.3bz: The Signaling Physics of Multi-Gigabit Ethernet

Traditional 1 GbE and 10 GbE use NRZ (Non-Return-to-Zero) signaling, where a single electrical pulse represents a single bit — $0\\, \\text{V}$ for logical 0 and a positive voltage for logical 1. This is a binary modulation scheme with a spectral efficiency of $1\\, \\text{bit/s/Hz}$. To achieve 10 Gbps over a Category 6A copper cable, the signal must operate at $2,500\\, \\text{Mbaud}$ (megabaud). As speeds increase past 25 Gbps, NRZ signaling becomes impractical due to the FEXT (Far-End Cross-Talk) and Insertion Loss constraints of twisted-pair copper — the signal attenuates below the noise floor before the receiver can sample it.

PAM-4 (Pulse Amplitude Modulation with 4 levels) solves this by encoding two bits per symbol. Instead of two voltage levels, PAM-4 uses four voltage levels: $-1\\, \\text{V}$ (00), $-0.33\\, \\text{V}$ (01), $+0.33\\, \\text{V}$ (10), and $+1\\, \\text{V}$ (11). This doubles the spectral efficiency: a PAM-4 link operating at $2,500\\, \\text{Mbaud}$ delivers $5\\, \\text{Gbps}$ over the same cable plant that would only support $2.5\\, \\text{Gbps}$ with NRZ.

IEEE 802.3bz (2.5 GbE and 5 GbE) and 802.3by (25 GbE / 100 GbE) both adopt PAM-4 for the electrical sub-layer. The trade-off is Signal-to-Noise Ratio (SNR): the voltage gap between adjacent PAM-4 levels is only one-third that of NRZ, requiring a 9.5 dB higher SNR for the same bit error rate (BER). To compensate, 802.3bz mandates LDPC (Low-Density Parity-Check) forward error correction at the physical coding sub-layer (PCS). LDPC adds parity overhead of roughly 6.7% but reduces the effective BER from $10^{-6}$ to $10^{-12}$, meeting the Ethernet standard of one error per $10^{12}$ bits transmitted.

The Auto-Negotiation Handshake in Multi-Gigabit Ethernet

For PAM-4 to work, both link partners must agree on the modulation scheme, the number of lanes, and the FEC mode. This is negotiated during the Auto-Negotiation (AN) phase defined in Clause 73 of IEEE 802.3. AN uses a base page exchange of 16 bits, with next-page messages to advertise PAM-4 capability. The negotiation sequence is:

1. Both ports transmit FLP (Fast Link Pulse) bursts containing their capabilities.
2. The receiver decodes the base page and selects the highest common speed (e.g., if one port supports 10G and 2.5G and the other supports 5G and 2.5G, they select 2.5G).
3. If both advertise PAM-4 ability, the link transitions to PAM-4 training mode.
4. During training, a known pseudo-random bit sequence (PRBS-13) is transmitted for 200–500 ms while the receiver adapts its equalization filter taps (Decision Feedback Equalizer, DFE) to cancel channel ISI (Inter-Symbol Interference).
5. Training completes, FEC is enabled, and the link transitions to the data mode.

The equalization training phase is critical: without DFE, the residual ISI from PAM-4 encoding would produce a vertical eye closure of $> 80\\%$ at the receiver, making reliable bit recovery impossible. The DFE adaptively subtracts post-cursor ISI using a tapped delay line with 10–14 taps, each tap coefficient being a 6-bit signed integer updated via a least-mean-squares (LMS) algorithm.

7. Energy-Efficient Ethernet (802.3az): The LPI Sleep-State Trade-Off

Most network ports consume nearly as much power when idle as when transmitting at full speed. In a data center with 10,000 switch ports, this idle power accounts for approximately 60% of total networking energy expenditure. IEEE 802.3az (Energy-Efficient Ethernet) addresses this by introducing a Low Power Idle (LPI) mode at the physical layer.

The LPI Duty Cycle

In LPI mode, the transmitter effectively shuts off the analog front-end for a period of Ts (Sleep Time), then briefly transmits a refresh signal to maintain link synchronization. The duty cycle is governed by:

For 10 GbE twisted-pair, a typical LPI cycle is $T_{\\text{sleep}} = 20\\, \\mu\\text{s}$, $T_{\\text{refresh}} = 2\\, \\mu\\text{s}$, and $T_{\\text{wake}} = 4\\, \\mu\\text{s}$. At 99% link idle, the power savings reach approximately 70%, dropping from 2 W per port to 0.6 W. However, the trade-off is Wake-up Latency: when a burst of traffic arrives during LPI sleep, the transmitter must flush the PCS buffer, re-synchronize the clock data recovery (CDR) circuit, and re-train the equalizer — a process that adds 12–30 microseconds of added latency to the first packet of every idle-to-active transition.

The Buffer Flush Problem in Lossless Fabrics

In a lossless RoCEv2 or FC-over-Ethernet fabric, EEE is a liability. LPI sleep causes the receiver's input buffer to drain because no data is arriving, and when the burst arrives after a wake transition, the headroom in the lossless buffer is insufficient to absorb the micro-burst — leading to PFC (Priority Flow Control) pause frames being generated. The pause frames propagate upstream, creating a PFC Storm that can spread congestion across the entire fabric. Because of this, EEE is universally disabled on data center switches carrying storage or HPC traffic, despite the 2–3 W per-port power savings.

Conclusion: The Foundation of Connectivity

The Ethernet frame has survived since 1980 because it is simple, efficient, and extensible. While we now have fiber optic cables capable of $400\\,\\\\text{Gbps}$, the fundamental structure of the Ethernet II frame remains identical to the one designed by Bob Metcalfe decades ago. It is the language that the world's hardware speaks.

Frequently Asked Questions