If you work in networking, you know the drill: optimise for throughput, minimise packet loss, and let the upper layers deal with the rest. That approach has worked well for decades.
However, AI workloads place very different demands on networks than traditional enterprise applications.
When thousands of GPUs communicate with each other to train a single AI model, they generate enormous amounts of east-west traffic and frequent bursts of data exchange. These workloads have exposed limitations in traditional data centre networking designs and are driving the industry towards new approaches.
Two technologies are at the centre of this shift: Ultra Ethernet and 1.6T Ethernet.
Although they are often discussed together, they solve different problems:
- 1.6T Ethernet is about moving more data through the network.
- Ultra Ethernet is about moving AI traffic more efficiently.
Think of 1.6T as building wider motorways, while Ultra Ethernet is about managing traffic more intelligently.
If you're a network operator trying to make sense of the latest acronyms, here's a plain-English introduction to what these technologies are, how they work, and why they matter.
1. What is 1.6T Ethernet? (The physical speed)
Simply put, 1.6T stands for 1.6 Terabits per second. It is the next major Ethernet speed generation after 800G.
To put that into perspective, a 1.6T link can move approximately 200 gigabytes of data every second. That's roughly 1,600 times faster than the 1 Gigabit Ethernet connection still found on many laptops and office devices.
The technical specifications are being defined by the IEEE 802.3dj task force, which is developing standards for 200G, 400G, 800G and 1.6T Ethernet.
How does 1.6T work?
Rather than creating a single massive data channel, manufacturers achieve 1.6T by combining multiple high-speed lanes.
224G lanes
A typical 1.6T connection uses eight parallel lanes operating at approximately 200–224 Gigabits per second each.
8 × 200G = 1.6T
PAM4 signalling
To transmit data at these speeds, modern Ethernet uses PAM4 (Pulse Amplitude Modulation with four levels).
Traditional signalling uses two voltage levels to represent binary values. PAM4 uses four distinct signal levels, allowing each symbol to carry twice as much information and effectively doubling data density.
The real-world challenge: physics
At these speeds, electrical signals degrade rapidly.
Running 200G+ lanes over copper becomes increasingly difficult as distance increases. Signal loss, interference and power consumption all become significant engineering challenges.
As a result, hyperscale data centres are increasingly adopting:
- Active Electrical Cables (AECs), which include signal-conditioning electronics inside the cable assembly
- Active Optical Cables (AOCs)
- High-density optical interconnects such as MPO-based fibre systems
The higher the bandwidth, the more important optical connectivity becomes.
2. What is Ultra Ethernet? (The smart brain)
Having a 1.6T link is only useful if the network can efficiently handle the traffic flowing through it.
Historically, many AI clusters have used RoCEv2 (RDMA over Converged Ethernet) to achieve low-latency communication between GPUs.
RoCEv2 performs best on highly lossless networks. Even small amounts of packet loss can impact performance because distributed AI workloads are extremely sensitive to latency and retransmissions.
To understand why this matters, imagine 10,000 GPUs training a large language model. Every few milliseconds, they exchange updated model parameters with thousands of other GPUs. If some GPUs are delayed waiting for network traffic, the entire training job can slow down while expensive accelerators sit idle.
To address these challenges, companies including AMD, Broadcom, Cisco, Intel, Meta and Microsoft formed the Ultra Ethernet Consortium (UEC). Their goal is to make Ethernet better suited for large-scale AI infrastructure.
Rather than relying on traditional networking approaches, Ultra Ethernet introduces new mechanisms designed specifically for AI traffic patterns.
Adaptive load balancing
Instead of forcing traffic through a single path, Ultra Ethernet can distribute traffic across multiple available routes, reducing hotspots and improving bandwidth utilisation.
Out-of-order delivery
Traditional networking often expects packets to arrive in sequence.
Ultra Ethernet allows packets to take different paths through the network and arrive out of order, with the receiving system responsible for reassembling them correctly.
Smarter recovery mechanisms
If packets are lost due to congestion or other issues, Ultra Ethernet is designed to recover efficiently without causing widespread performance degradation across the cluster.
The goal is simple: keep GPUs busy and minimise time spent waiting on the network.
3. The comparison: InfiniBand vs Ultra Ethernet
Today, large-scale AI networking is largely split between two approaches.
On one side is InfiniBand, which has long been the preferred technology for high-performance computing and AI supercomputers.
InfiniBand is largely controlled by Nvidia and offers a tightly integrated ecosystem of adapters, switches and management software. It delivers extremely low latency and excellent performance at scale.
On the other side is Ultra Ethernet.
Rather than creating a separate networking technology, Ultra Ethernet builds on the Ethernet ecosystem that already dominates modern data centres.
The key advantage is openness.
Because Ethernet is supported by multiple hardware vendors, operators can potentially build AI infrastructure using components from companies such as Broadcom, Intel, AMD, Cisco, Arista and others.
For hyperscalers, that flexibility can help reduce costs, increase supplier choice and minimise dependence on a single vendor ecosystem.
4. Is Ultra Ethernet relevant beyond AI datacenters?
Even if you're not building AI superclusters, these developments are worth watching.
AI infrastructure is changing how modern data centres are designed.
From a network operations perspective:
- East-west traffic volumes continue to grow
- Congestion becomes increasingly expensive
- Network visibility becomes more important as clusters scale
- High-speed optical interconnects are becoming mainstream
Many of the innovations being developed for AI environments will eventually influence broader data centre networking designs.
5. When is this actually happening?
The transition is already underway.
Today
Ultra Ethernet 1.0 has been released, and major vendors are building products around its specifications.
Meanwhile, 800G Ethernet remains the primary deployment speed for most large-scale data centre networks.
The next few years
Hyperscalers are expected to begin adopting 1.6T deployments for new AI-focused infrastructure as hardware availability increases and standards mature.
However, 800G is likely to remain the dominant deployment speed across much of the industry for some time.
Long term
Most enterprises, campuses and office networks have no immediate need for 1.6T connectivity.
For the foreseeable future, technologies such as 10G, 25G, 100G and 400G Ethernet will continue to meet the needs of the majority of organisations.
Ethernet is evolving faster than ever
Ethernet used to evolve in steady cycles. New speeds arrived every several years, and each generation took time to move from hyperscale deployments into wider use.
That pattern is changing. AI infrastructure is accelerating the roadmap. The industry has moved from 400G to 800G and is already heading toward 1.6T in just a few years. At the same time, Ultra Ethernet is being developed to address not just speed, but how traffic behaves inside large AI clusters.
This is one of the fastest periods of change in Ethernet’s history. What used to take close to a decade is now happening in two-to-three year cycles at the high end of the market.
Also, it is noteworthy to mention that Ultra Ethernet and 1.6T are not isolated upgrades. They are signs of a broader shift where AI is now a major driver of network design, from link speeds to congestion handling and traffic flow.
