Ethernet Evolution: From ALOHA to 100 Gbps

In 1980, 10 Mbps Ethernet was revolutionary. Today, 400 Gbps Ethernet is deployed in hyperscale data centers, with 800 Gbps standards in development and 1.6 Tbps on the horizon. This represents a 40,000-fold increase in speed while maintaining fundamental protocol compatibility—an engineering achievement unparalleled in computing history.

To put this in perspective: if automobile speeds had improved proportionally since 1980, cars would now travel at over 4 million miles per hour—fast enough to reach the moon in under four minutes.

This page traces Ethernet's speed evolution, examining the engineering breakthroughs that enabled each generational leap and the design decisions that maintained compatibility across this remarkable progression.

For fifteen years, 10 Mbps was the only Ethernet speed. This long period wasn't stagnation—it was consolidation. During this time, Ethernet transitioned from thick coaxial cable to thin coaxial to twisted pair to fiber, establishing the physical layer flexibility that would enable future speed increases.

The 10 Mbps ecosystem:

Why 10 Mbps worked for 15 years:

In the 1980s and early 1990s, 10 Mbps was more than adequate for typical office applications:

• Text files: A 50 KB document transfers in 40 milliseconds
• Simple graphics: A 500 KB image transfers in 400 milliseconds
• Printing: Laser printers of the era couldn't render faster than 10 Mbps delivery

Network utilization rarely exceeded 30% in typical office environments. Shared 10 Mbps was sufficient for dozens of users.

The pressure for more speed:

By the mid-1990s, several factors created demand for faster Ethernet:

1. Graphical applications: Windows, Mac GUIs, and emerging web browsers used larger files
2. Client-server computing: Databases and file servers became bottlenecks
3. Multimedia: Audio and video applications required sustained bandwidth
4. Internet growth: Corporate internet connections demanded more backbone capacity
5. Server consolidation: Centralizing servers created traffic concentration points

The IEEE 802.3u standard, ratified in 1995, defined Fast Ethernet at 100 Mbps. This 10x speed increase presented significant engineering challenges while maintaining the fundamental Ethernet protocol.

The core challenge: CSMA/CD timing

Recall that Ethernet's minimum frame size (64 bytes) relates directly to the collision detection requirement. At 10 Mbps:

• Slot time = 512 bit times = 51.2 μs
• Round-trip propagation at speed of light ≈ 25.6 μs for 2.5 km
• This allows collision detection across maximum network diameter

At 100 Mbps, the same 512 bit times = 5.12 μs. At the speed of light, signals travel only about 250 meters in this time—meaning the maximum collision domain shrinks to roughly 200 meters (accounting for cable and electronic delays).

Fast Ethernet standards:

Encoding schemes:

Fast Ethernet couldn't simply transmit at 100 MHz on copper—the electrical characteristics of twisted pair cables would cause unacceptable signal degradation. Instead, sophisticated encoding schemes were developed:

100BASE-TX (MLT-3 encoding):

1. Data is first encoded using 4B5B (4 bits to 5 bits), providing clock recovery
2. The 5-bit output is then scrambled to prevent DC bias
3. Finally, MLT-3 (Multi-Level Transmit) encoding uses three voltage levels (+V, 0, -V)
4. This reduces the fundamental frequency to 31.25 MHz despite 125 Mbaud signaling

100BASE-T4 (8B6T encoding):

1. Designed for existing Cat 3 wiring
2. Uses all four pairs (vs. two for TX)
3. Three pairs transmit, one receives (or vice versa)
4. 8B6T encoding converts 8 bits to 6 ternary symbols
5. Each pair runs at only 25 MHz, within Cat 3 capability

These clever encoding techniques allowed 100 Mbps over cables designed for 10 Mbps operation.

While Fast Ethernet increased raw speed 10x, a parallel development would effectively double network capacity again: full-duplex operation.

Half-duplex vs. Full-duplex:

The switch transformation:

Full-duplex Ethernet required abandoning hubs for switches. Unlike hubs, which repeat signals to all ports, switches:

1. Learn MAC addresses — Build a table mapping addresses to ports
2. Forward selectively — Send frames only to the destination port
3. Buffer frames — Store frames when output ports are busy
4. Enable full-duplex — Create dedicated send/receive paths per port

With switches, each station has a dedicated 100 Mbps channel to the switch. Two stations can exchange data at 100 Mbps in each direction simultaneously—effectively 200 Mbps of throughput per station pair.

Flow control in full-duplex: IEEE 802.3x PAUSE

With CSMA/CD's natural backpressure removed, full-duplex networks needed an alternative congestion control mechanism. IEEE 802.3x (1997) introduced the PAUSE frame:

1. When a receiver's buffers fill, it sends a PAUSE frame to the sender
2. The PAUSE frame specifies a wait time (in 512-bit units)
3. The sender must stop transmitting for that duration
4. After the pause, or if a PAUSE with time=0 arrives, transmission resumes

PAUSE is a Layer 2 mechanism—it operates below TCP/IP and can stop traffic instantly. However, it's a blunt instrument: it stops all traffic on a link, regardless of destination or priority. This limitation led to later developments like Priority Flow Control (PFC) for data center applications.

IEEE 802.3z, ratified in 1998, defined Gigabit Ethernet (GbE) at 1000 Mbps. This represented another 10x speed increase and another set of engineering challenges.

The half-duplex challenge:

At 1000 Mbps, the 512-bit slot time shrinks to just 512 nanoseconds. Light travels only about 100 meters in this time—making even a single cable run from station to switch problematic for CSMA/CD.

The IEEE's solution was controversial: carrier extension. For half-duplex Gigabit Ethernet, the slot time was increased to 4096 bits (512 bytes), allowing approximately 200-meter collision domains. Short frames (under 512 bytes) were padded with carrier extension symbols—essentially meaningless signals that maintained transmission long enough for collision detection.

This was inefficient for small frames:

Frame bursting:

To mitigate carrier extension's inefficiency, frame bursting was also introduced. A station with multiple frames to send could transmit them in a burst, with only the first frame requiring carrier extension. Subsequent frames in the burst could use minimal inter-frame gaps, amortizing the extension overhead across multiple frames.

The practical reality: Full-duplex dominated

In practice, half-duplex Gigabit Ethernet was rarely deployed. By 1998, switches were affordable, and the performance advantages of full-duplex were overwhelming. Half-duplex GbE's carrier extension and frame bursting were implemented mainly for standards compliance.

Gigabit Ethernet standards:

1000BASE-T: Engineering marvel

1000BASE-T (IEEE 802.3ab, 1999) achieved Gigabit speeds over standard Cat 5e twisted pair—the same cabling as 100BASE-TX. This required extraordinary signal processing:

1. All four pairs used — Each pair carries data simultaneously in both directions
2. PAM-5 encoding — 5 voltage levels (-2, -1, 0, +1, +2) encode 2 bits per symbol
3. 4D-PAM5 at 125 Mbaud — 4 pairs × 2 bits × 125 M symbols = 1000 Mbps
4. Echo cancellation — DSP removes echo from simultaneous transmit/receive on same pair
5. Near-end crosstalk cancellation — DSP removes interference between pairs

The transceivers contained more processing power than entire 1980s computers. Yet the standard maintained plug-compatibility with existing cabling infrastructure.

IEEE 802.3ae, ratified in 2002, defined 10 Gigabit Ethernet (10GbE). At this speed, supporting CSMA/CD would limit collision domains to under 20 meters—clearly impractical. Accordingly, 10GbE was defined as full-duplex only, formally ending CSMA/CD's role in Ethernet.

10GbE: Two distinct markets

10GbE standards addressed two different application domains:

10 Gigabit Ethernet standards:

The rise of 40G and 100G:

IEEE 802.3ba (2010) defined 40 and 100 Gigabit Ethernet. Rather than increasing per-lane speeds beyond manufacturing limits, these standards used parallel lanes:

• 40GBASE-SR4: 4 × 10 Gbps lanes over parallel fiber
• 100GBASE-SR10: 10 × 10 Gbps lanes (later replaced by 4 × 25G)
• 100GBASE-LR4: 4 × 25 Gbps wavelengths on single fiber (WDM)

Higher speeds: 200G, 400G, and beyond

IEEE 802.3bs (2017) defined 200 and 400 Gigabit Ethernet using higher lane rates (50 Gbps, then 100 Gbps per lane):

• 200GBASE-SR4: 4 × 50 Gbps
• 400GBASE-SR16: 16 × 25 Gbps (later replaced by 8 × 50G or 4 × 100G)
• 400GBASE-DR4: 4 × 100 Gbps over parallel single-mode fiber

IEEE 802.3db (2022) defined 800 Gigabit Ethernet, and work on 1.6 Terabit Ethernet is underway.

Let's consolidate Ethernet's speed evolution with a comprehensive timeline:

The doubling period:

Observe that Ethernet speed has roughly doubled every 2-3 years, consistent with Moore's Law for the underlying semiconductor technology:

• 1983-1995: 10 Mbps (12 years of consolidation)
• 1995: 100 Mbps (10x)
• 1998-1999: 1 Gbps (10x)
• 2002: 10 Gbps (10x)
• 2010: 100 Gbps (10x)
• 2022: 800 Gbps (8x since 2010)

Each generational leap required innovations in encoding, optics, signal processing, and manufacturing. Yet through all these changes, the fundamental Ethernet frame format remains compatible with the 1982 Ethernet II specification.

Each speed increase required more sophisticated encoding to transmit higher bit rates over physical media. The progression of encoding schemes tells the story of increasing engineering complexity:

Why encoding efficiency matters:

High-efficiency encoding is critical at higher speeds:

• 10BASE-T (Manchester): 50% efficiency means 20 MHz signaling for 10 Mbps—manageable
• 10GBASE-T (65B/66B + PAM-16): 97% efficiency means ~400 MHz signaling for 10 Gbps—at the edge of Cat 6a capability

At 100 Gbps and beyond, every percent of encoding overhead translates to billions of additional symbols per second—requiring faster electronics, more power, and higher manufacturing precision.

The PAM-4 revolution:

For 25+ Gbps lanes, PAM-4 (Pulse Amplitude Modulation with 4 levels) became essential:

• 2 bits encoded per symbol (vs. 1 bit for NRZ/PAM-2)
• Halves the required baud rate for a given bit rate
• 50 Gbps link uses 26.5625 Gbaud (vs. 53.125 Gbaud for NRZ)
• Enables higher speeds without proportionally faster electronics

The trade-off: PAM-4's closer voltage levels require better signal-to-noise ratio, more sophisticated equalization, and more powerful forward error correction (FEC).

We've traced Ethernet's remarkable speed evolution from 10 Mbps to 800 Gbps and beyond. Let's consolidate the key insights:

What's next:

Having surveyed the speed evolution, we'll now examine specific milestones in detail: the jump from 10 Mbps to 100 Gbps. The next page explores the engineering challenges and solutions at each major speed tier.