Ultra Ethernet Consortium Announces the Release of UEC Specification 1.0

[Nomad76]

The Ultra Ethernet Consortium (UEC) today announces the release of UEC Specification 1.0, a comprehensive, Ethernet-based communication stack engineered to meet the demanding needs of modern Artificial Intelligence (AI) and High-Performance Computing (HPC) workloads. This release marks a critical step forward in redefining Ethernet for next-generation, data-intensive infrastructure.

UEC Specification 1.0 delivers a high-performance, scalable, and interoperable solution across all layers of the networking stack—including NICs, switches, optics, and cables—enabling seamless multivendor integration and accelerating innovation throughout the ecosystem.



The UEC specification is advancing industry-wide adoption by promoting open, interoperable standards that prevent vendor lock-in. With active implementations and compliance programs already in progress, UEC is paving the way for a unified and accessible ecosystem across the sector.

"The Ultra Ethernet 1.0 Specification is the result of an outstanding collaboration of AI, HPC, and networking experts, system and silicon vendors, and network operators," added Hugh Holbrook, Chair of the Technical Advisory Committee, Ultra Ethernet Consortium. "It incorporates a wealth of knowledge, experience, and ideas related to applications, transport protocols, congestion control, direct memory access, Ethernet link and PHY technologies, and network security. This is a landmark development for the industry, and I'm thrilled for the public release of Ultra Ethernet to further boost Ethernet's success in AI and HPC."

Addressing Critical Infrastructure Needs
As AI and HPC workloads continue to evolve at unprecedented speed, UEC's purpose-built Ethernet innovation delivers:

• Modern RDMA for Ethernet and IP - Supporting intelligent, low-latency transport for high-throughput environments.
• Open Standards and Interoperability - Avoids vendor lock-in while accelerating ecosystem-wide innovation.
• End-to-End Scalability - From routing and provisioning to operations and testing, UEC scales to millions of endpoints.

Designed for Real-World Integration
Built on the globally adopted Ethernet standard, UEC 1.0 simplifies deployment across the full technology stack—from hardware to applications. Its familiarity, combined with robust tooling, makes it especially valuable to cloud infrastructure operators, hyperscalers, DevOps teams, and AI engineering leads seeking low-friction adoption paths.

"As Chair of the Ultra Ethernet Consortium, I'm proud to announce the release of the UEC 1.0 specification—a major milestone in our mission to redefine Ethernet for the age of AI and high-performance computing," said Dr. J Metz, Chair of the Steering Committee, Ultra Ethernet Consortium. "This standard is the result of unprecedented collaboration across the industry, and it delivers the low-latency, high-bandwidth, and intelligent transport needed for the most demanding workloads of today and tomorrow. The UEC 1.0 release is just the beginning—Ultra Ethernet is here, and it's built to scale the future."

View at TechPowerUp Main Site | Source

 

[Shihab]

So much text to just say "F u, Nvidia!"

 

[Juventas]

So you can do 100 Gbps now with CAT8 cables. I wonder why using cheap UTP cables is important when you need switches that cost over $10k and beyond. We already have standard SFP and copper DAC cables that can handle up to 400 Gbps.

 

[Dristun]

Another extra $50 for a new motherboard!

 

[lexluthermiester]

So much text to just say "F u, Nvidia!"

Wait up, THAT is what you get from this article?!? Are you serious?

When is this brand bashing crap going to stop?

 

[AnarchoPrimitiv]

Wait up, THAT is what you get from this article?!? Are you serious?

When is this brand bashing crap going to stop?

Those poor, poor multi-billion dollar companies with basically infinite power and zero accountability....what ever would they do if there weren't people like yourself looking out for underdogs like Nvidia, right? Despite the fact that 99.99% of the "bad treatment" is FROM corporations and DIRECTED at consumers and stake holders (not shareholders), the "final straw" for you is the rare instance when the treatment goes in the opposite direction? Interesting.

 

[Juventas]

Wait up, THAT is what you get from this article?!? Are you serious?

When is this brand bashing crap going to stop?

Nvidia is a player in data centre NICs, but yeah, they don't have the dominance like with GPUs. I believe Broadcom and Intel are the top two.

 

[R-T-B]

Wait up, THAT is what you get from this article?!? Are you serious?

When is this brand bashing crap going to stop?

It's a pretty direct competitor to NVLink's dominance in AI is I think his angle. Either that or NIC stuff.

 

[lexluthermiester]

Nvidia is a player in data centre NICs, but yeah, they don't have the dominance like with GPUs. I believe Broadcom and Intel are the top two.

But NVidia? Seriously?...

It's a pretty direct competitor to NVLink's dominance in AI is I think his angle. Either that or NIC stuff.

Ok? That's more than a bit of a stretch, but whatever.. I think they just wanted to moan a bit of brand bashing.

 

[SRB151]

But NVidia? Seriously?...

Ok? That more than a bit of a stretch, but whatever.. I think they just wanted to moan a bit of brand bashing.

No, the reason AMD doesn't scale past a single server is that NVlink is proprietary. Ultra Ethernet is the open source answer to NVlink. It will be used on AMD's MI400 series.
Their dominance is pretty much due to CUDA and NVlink. With these out of the way, the door will be open for more (read: cheaper) options for AI.

 

[igormp]

No, the reason AMD doesn't scale past a single server is that NVlink is proprietary. Ultra Ethernet is the open source answer to NVlink. It will be used on AMD's MI400 series.
Their dominance is pretty much due to CUDA and NVlink. With these out of the way, the door will be open for more (read: cheaper) options for AI.

NVLink as a inter-node setup is a really recent thing. Before Hopper all clusters were either using infiniband or just Ethernet, and NVLink was solely a intra-node thing.
I guess you meant to say NVSwitch and not NVLink. The competitor to NVLink is UALink, which has been delayed for quite a bit, iirc due to some broadcom issues.

In the same way Nvidia has bought Mellanox in order to have some strong networking in their portfolio, AMD has bought Xilinx and Pensando, both of which have their own very capable networking solutions, which already included 200G devices, and 400G as of last year.
The fact that AMD can't do proper scale out is their own fault, and has nothing to do with NVLink or CUDA.