AI training turns communication into the bottleneck — which is exactly the problem both 1.6T Ethernet and Ultra Ethernet are racing to solve

1. What Is Ultra Ethernet? Ethernet's Answer to InfiniBand

Ultra Ethernet is a specification and open ecosystem driven by the Ultra Ethernet Consortium (UEC), founded in 2023 by a group of silicon, switch, and hyperscaler companies — including AMD, Broadcom, Cisco, Intel, Meta, and Microsoft. Its stated goal is simple: make the Ethernet everyone already runs good enough for AI and HPC scale-out, so buyers have an open alternative to Nvidia's InfiniBand.

To understand why it exists, you have to look at the problem it targets. Distributed AI training is dominated by collective communication — every few milliseconds, each of thousands of GPUs exchanges gradients and parameters with thousands of others. When even a few GPUs wait on the network, the whole training joba slows down, and expensive accelerators sit idle. That wait time — not raw bandwidth — is what Ultra Ethernet attacks.

2. The Problem: Why AI Clusters Stall on RoCEv2

Most AI clusters today run RoCEv2 (RDMA over Converged Ethernet) for GPU-to-GPU communication. RoCEv2 works beautifully on a perfectly lossless network — and punishes you brutally the moment a single packet drops.

In a 10,000-GPU training run, a dropped packet in one flow can stall the entire AllReduce collective, because every GPU has to wait for the slowest one. Traditional Ethernet's answers to this are blunt instruments:

• Priority Flow Control (PFC) stops traffic to prevent loss — but can cause head-of-line blocking and even deadlocks across a large fabric.
• Single-path routing wastes available bandwidth and creates hotspots on congested links.
• In-order delivery forces packets to arrive in sequence, so one slow packet holds up everything behind it.

The result is what AI operators actually measure and dread: inflated tail latency and longer job completion time (JCT). Ultra Ethernet exists to replace those blunt instruments with smarter ones.

3. How Ultra Ethernet Works: Spraying, Out-of-Order & Smarter Recovery

The UEC specification (1.0 released in 2025) redesigns the transport layer around how AI traffic actually behaves. Three mechanisms do most of the work:

On top of these sit UEC's other building blocks: congestion telemetry for visibility, in-network collective operations that offload parts of AllReduce into the switch, and a defined Ultra Ethernet Transport (UET) protocol. Together they are an attempt to approach InfiniBand's latency and lossless behavior while staying in the open Ethernet world.

Packet spraying and out-of-order delivery are the two ideas that separate Ultra Ethernet from the RoCEv2 status quo

4. What Is 1.6T Ethernet? IEEE 802.3dj & the 200G Lane

While Ultra Ethernet reworks the traffic layer, 1.6T Ethernet reworks the road itself. It is the next rung on the speed ladder — 1.6 terabits per second, roughly 200 gigabytes per second, or about 1,600 times a standard 1 Gigabit link — being defined by the IEEE 802.3dj task force.

The architectural breakthrough is the 200 Gbps serial lane. Instead of bonding sixteen 100G lanes, a 1.6T link typically runs 8 lanes of 200G each (8 x 200G), built on 224G SerDes and PAM4 signaling. That lane-rate doubling is what makes the next generation of optics and switches possible — and it is why 400G, 800G, and 1.6T are all being specified together in the same IEEE effort.

For the cabling and optics implications of moving from 400G to 800G to 1.6T, see our 800G/1.6T data center cabling trends.

5. Ultra Ethernet vs 1.6T: Transport vs Physical Layer

It helps to keep a single mental model. Think of 1.6T Ethernet as building a wider highway, and Ultra Ethernet as installing smarter traffic management. You can have one without the other:

• 1.6T without Ultra Ethernet: more raw capacity, but the same congestion, packet-loss, and tail-latency problems — just happening faster.
• Ultra Ethernet without 1.6T: you can run UEC transport over 400G or 800G links today and still cut job completion time by using bandwidth more efficiently.

The two are complementary layers, and the AI industry is adopting both in parallel. 1.6T supplies the raw bandwidth a full-bisection AI fabric demands; Ultra Ethernet ensures that bandwidth is actually used well. This is the deeper shift the industry is signaling: AI is now the primary driver of network design, from link speed to congestion control.

6. The 1.6T Hardware Picture: SerDes, PAM4 & Cabling

Making 1.6T work in silicon and copper is where the real engineering lives. A few pieces to know:

• 224G SerDes with PAM4: each lane encodes two bits per symbol using four-level signaling, doubling data density — at the cost of tighter signal-integrity budgets. Receivers now use heavy DSP and forward-error correction (RS-FEC) to recover signals through lossy channels.
• Copper has hit its distance wall: at 200G per lane, passive DAC reaches only about two meters. Beyond that, deployments move to AEC (active electrical cables) to ~7 m, then AOC (active optical cables) and pluggable optics over fiber.
• Optics and form factors: the leading 1.6T module form factor is OSFP-XD, packing 16 electrical lanes. Longer-reach links lean on single-mode OS2 fiber, which scales cleanly to 200G-per-lane over distance in a way multimode cannot.

At 200G per lane, the physical layer shifts decisively toward single-mode fiber and high-density MPO interconnects

For how MPO fiber scales to these speeds, see our guide on MPO fiber for 400G/800G links.

7. Timeline & What It Means for Your Network

The honest deployment picture as of 2026:

• 800G is the workhorse. It is the dominant deployed speed in hyperscale AI fabrics, and the natural upgrade target for most large-scale builds today.
• 1.6T is coming, not yet ubiquitous. Early hyperscalers are sampling 1.6T silicon and optics, with broader adoption expected as the 802.3dj standard finalizes and the supply chain matures.
• Ultra Ethernet is already shipping into designs. The UEC 1.0 spec is out, and major silicon and switch vendors are building compliant products around it.
• Most enterprises can wait. Campus, office, and general data center networks are well served by 10G, 25G, 100G, and 400G — none of this AI-scale plumbing changes that.

For the deeper dive into how AI is reshaping the physical cabling plant, see our structured cabling for AI data centers.

Key Questions (FAQ)

AMPCOM Technical Team

Industry experts with 17+ years in enterprise network infrastructure and structured cabling systems