Name: Overview of Ultra Ethernet Presented by UEC
Uploaded: 2026-03-28T11:07:38.000Z
Duration: 15 min 53 s
Description: Thousands of YouTube videos with English-Chinese subtitles! Now you can learn to understand native speakers, expand your vocabulary, and improve your pronunciation...

By the way, this is going to be an audience participation portion of the day.

So I'm going to ask things and you're going to laugh.

I am the chair of the Ultra Ethernet Consortium, and

I think I may have the wrong slides up here, but we'll go make it through.

But my colleague Uri and I did a presentation yesterday on the Ultra Ethernet, and

I'm going to be talking to you a little bit about today.

So have you gotten stickers for Ultra Ethernet?

If I had a sticker on my laptop for every time someone offered me a sticker, I'd have a 10-pound laptop.

And it's a Mac, so it's already 9 pounds.

So what we're going to be talking about today is an introduction of the Ultra Ethernet Consortium, and

we're going to talk a little bit about why we're doing what we're doing and what we're doing what we're doing.

So if you've already been familiar with this, you've heard them talk about Ultra Ethernet, the UEC Consortium, and

It's a big deal for a lot of different companies.

As a matter of fact, there's more than 100 different companies.

As the fastest-growing organization in the Linux Foundation history, we've gone from 10 companies to 100 companies in less than a year.

We have a lot of different companies who are looking to try to provide an Ethernet-based open ecosystem for AI and HPC workloads.

That sounds like a lot of marketing garbage.

But in reality he's not even saying, "Yep.

Yes, it does." Sounds like marketing garbage.

One of the things that we like about Ethernet is the fact that it is incredibly flexible.

It is a great general-purpose network, right?

It does an awful lot of things for a lot of different types of workloads.

You can do a lot of different types of things with Ethernet.

The problem is that it's so generalized that if you have a specific workload, making sure that each and every one of those layers is tuned is something of a problem, right?

If you want to get the physical layer to connect to the link layer, if you want to get the link knee bone connects to the hip bone, right?

If you want to do that, it actually takes quite a bit of work.

And so what we at the Ultra Ethernet Consortium are trying to do is make sure that that synchronization that goes across all of the different layers is tuned to a particular workload.

So we have taken the standard Ethernet from the physical layer up into the transport layer and above, and

we've made some tweaks up and down the stack to be able to handle different types of workloads, including AI and HPC.

Now, AI and HPC have some similarities, but they're not 100% completely the same.

And when you start to dig deep into the way these workloads actually function, you start to realize that you need to have different types of approaches to having reliability, orderability, flexibility, multipathing, and

So we need to be able to tweak these things up and down the stack.

So that's what we're doing at Ultra Ethernet.

At the same time, you want to make sure that it actually works up and down the stack.

So we created four additional workgroups that are designed to get the checks and balances and the accountability for this.

And we've got compliance, testing, performance, debug, storage, management to make sure that all of these different horizontal layers are actually working in concert.

And as a result, we have, to my surprise, as the chair of Ultra Ethernet, found that we're not the only ones that want this to be solved, right?

We've got over 100 different companies now.

It just recently passed over the 100 mark.

And we've got over 1,400 people who have actually sat into the meetings and contributed.

It is an incredibly robust and diverse group of people from all over the planet.

So we've got people from the United States, from EMEA, from APAC.

So let's take a quick example of what it is that we're actually doing.

Generally speaking, when you have an enterprise data center and you're using Ethernet oh, I like the cameras.

So you have your general-purpose network, right?

But when you're talking about a specific type of workload, you want to make sure that the network actually winds up doing what you want it to do.

So you have backend networks for your AI and HPC workloads.

And then you have a third network, a third type of backend network, if you want to do kind of scale-up.

What's the difference between scale-up and scale-out?

Well, generally, in a very high-level view, the difference between scale-up and scale-out is that if I've got these accelerators, GPUs, and

I need to connect them together, I need to make sure that the memory regions for the GPUs all act as one huge honking GPU.

I got one GPU here, another GPU here, another GPU here.

And I'm going to scale this up so that it all looks like one GPU to do all the same work.

Subtitles ListPlay Video

Overview of Ultra Ethernet Presented by UEC

ultimately

specific

incredibly

incredible