Network

Packet Spraying & Flexible Order: The UEC's Secret Weapons for AI Networking

Traditional networks create traffic jams. Ultra Ethernet's secret weapons are packet spraying and flexible order—a new design that eliminates hotspots and unlocks the full power of the fabric by breaking all the old rules of packet delivery.

Arun Babu Neelicattu

06 Aug 2025 — 4 min read

Welcome back! In our last two posts, we established why the AI industry is desperately seeking an open, high-performance alternative to InfiniBand and introduced the Ultra Ethernet Consortium (UEC) as the vehicle for this revolution. We've talked a lot about the promise of a new, re-architected Ethernet, but now it's time to look under the hood.

What, exactly, makes the new Ultra Ethernet Transport (UET) protocol different? How does it solve the problems that crippled previous attempts like RoCE?

The answer lies in a pair of elegant, powerful, and deeply interconnected concepts that serve as the UEC's secret weapons: packet spraying and flexible delivery order. These aren't just incremental improvements; they represent a fundamental shift in how a network fabric operates, designed with the sole purpose of eliminating the bottlenecks that plague AI workloads.

ℹ️

This post is part of a series, you can see the complete series here.

The Root of All Evil: Hash Collisions and Idle Highways

To understand the solution, you first have to appreciate the problem. AI training generates massive, long-lived data transfers called "elephant flows." Imagine needing to move the entire contents of a library from one building to another—that's an elephant flow.

Traditional Ethernet networks use a technique called Equal-Cost Multi-Path (ECMP) to balance traffic. Think of it as a traffic dispatcher standing at a highway on-ramp with multiple lanes. The dispatcher looks at a car's origin and destination (the "flow") and assigns it to a single lane for its entire journey. This is done using a mathematical function called a "hash."

For normal, random internet traffic with millions of small cars going to different places, this works reasonably well. But AI traffic isn't random. It's a small number of giant elephant-sized trucks all trying to go between the same few buildings.

This leads to a disastrous problem called a hash collision. The dispatcher's hashing algorithm ends up assigning multiple elephant flows to the same lane. You get one highway lane completely gridlocked with giant trucks, while the other lanes sit nearly empty. This creates massive network imbalance, hotspots, and congestion, which is the primary cause of poor application performance.

Secret Weapon #1: Packet Spraying (No More Empty Lanes)

The UEC's solution to this is brilliantly simple: packet spraying.

Instead of assigning the entire truck (the flow) to one lane, UET breaks the truck's cargo down into thousands of individual boxes (packets). The network then "sprays" these packets across all available lanes between the source and destination.

It's a form of per-packet load balancing, not per-flow.

The result is near-perfect load balancing across the entire network fabric. Every path is utilized evenly. There are no more gridlocked lanes next to empty ones. By eliminating the network hotspots caused by hash collisions, packet spraying ensures that the full, aggregate bandwidth of the fabric is available to every application, all the time.

Secret Weapon #2: Flexible Order (The Magic Enabler)

Packet spraying is a fantastic idea, but it creates a new challenge. If you send all the boxes down different lanes at different times, they're going to arrive at the destination out of sequence.

Traditional networking protocols like TCP and even InfiniBand are built on a principle of strict in-order delivery. They are like an assembly line where every part must arrive in the exact right order, or the whole line stops. If packet #100 is delayed on one path, the receiver has to wait for it before it can process packet #101, even if #101 has already arrived on a faster path. This is the infamous "head-of-line blocking" problem that cripples RoCE.

UET solves this with its second secret weapon: flexible, out-of-order delivery.

UET fundamentally decouples the order of packets from the order of the final message. It says, "It doesn't matter what order the boxes arrive in, as long as we get all the boxes."

This is made possible by intelligent network interface cards (NICs) at the destination. The UET-compliant NIC receives the out-of-order stream of packets being sprayed across the fabric and is responsible for quickly and efficiently reassembling them into the correct message sequence before delivering the complete, finished data to the application's memory.

This is the true genius of the UEC design. Flexible order is the key that unlocks the immense power of packet spraying. It completely sidesteps head-of-line blocking. A delay, a bit of congestion, or a retransmission on one path has zero impact on the packets flying down other, healthier paths. This dramatically reduces "tail latency"—the time it takes for the very last packet of a message to arrive—which is the single biggest factor determining your overall Job Completion Time.

By combining these two weapons, the UEC has created a fabric that is inherently more resilient, more efficient, and far more predictable under the unique stress of AI workloads. It's a system designed not just to be fast in a perfect lab environment, but to be fast in the messy, chaotic reality of a multi-billion-dollar AI factory.

Next up: Meet the DPU: The "Third Pillar" of Compute That Makes AI Clouds Possible. We'll explore the unsung hero of the data center—the Data Processing Unit—and explain why offloading infrastructure tasks is non-negotiable for performance.