Supersizing Ethernet: Tech Vendors Are Looking To Lead the Way

The technology that powers enterprise networks and the internet will see some significant changes soon, with some of the biggest names in network hardware coming together. Cisco, Arista, HPE, Intel, Linux Foundation, and others have announced the Ultra Ethernet Consortium (UEC) to enhance and upgrade ethernet technology. 

A fundamental component of network architectures, the ethernet has existed for over 50 years. 

While it has witnessed remarkable changes over the decades, there’s a general agreement that it will need some performance enhancement to meet the needs of artificial intelligence (AI), which is gradually penetrating every part of the tech industry. AI's use cases are expanding dramatically from chatbots for public use to private research facilities. 

In this article, we’ll discuss the following:

• A deep dive into the announcement of the UEC.
• What the UEC entails for networking technologies.
• The need for significant hardware enhancement and overhauls in the current competitive AI-focused climate.

Not the article you’re looking for today? Try these out:

• The Future of Ethernet: What to Expect 
• Busting Myths About Third-Party Maintenance With Data
• Uptime Institute’s Five Data Center Predictions: Are They Coming True?

Introducing the Ultra Ethernet Consortium (UEC)

Arista’s CEO Jayshree Ullal announced the UEC in a blog post in July 2023, highlighting the need for terabit ethernet in an era of AI and machine learning. Besides Arista, the consortium’s founding members include AMD, Broadcom, Eviden, HPE, Intel, Meta, and Microsoft. The Linux Foundation hosts it. 

UEC’s mission is to enhance the capabilities of ethernet to support high-performance computing (HPC), so AI applications can be handled by networks readily. The consortium — along with IEEE, which has historically set the ethernet standards — is looking to take advantage of the broad adaptability and flexibility of ethernet to provide businesses and other entities a chance to leverage AI without being bogged down with infrastructural limitations. 

“As AI applications drive inference and training with massive compute processors (GPUs/CPUs/TPUs), one must reimagine the high-speed transit of mission-crucial workloads. The upshot is wire-rate delivery of large synchronized bursts of data using a familiar standards-based, Ethernet-based network,” Ullal explained in her blog.

UEC’s Ethernet Plans

In its white paper, published on the official website, the UEC provided a detailed overview of what it plans to achieve with the Ultra Ethernet Transport (UET) protocol. The paper states that improvements are needed to support the delivery of messages to all connected endpoints quickly and without delay. In other words, it plans to minimize tail latency. 

Here are the main capabilities and improvements UEC wants to add to the next-generation ethernet:

• Multi-pathing and packet spraying: Unlike the widely used equal cost multipath (ECMP) that uses a flow hash, Ultra Ethernet will employ packet spraying. This method will allow the use of all paths simultaneously, eliminating the need for load balancing. 
  
  
• Flexible ordering: With flexible ordering, the packets don’t need to be ordered before delivery, enabling collective communication operations such as AI, where large packets need to be delivered fast. 
  
  
• AI and HPC-optimized congestion control: Ultra Ethernet will employ modern congestion-controlling mechanisms to address the typical congestion points, like the link between the final switch and receiver. AI workloads are more likely to face congestion at the last link. A specific congestion control algorithm will work with packet spraying to reduce tail latency. 
  
  
• End-to-end telemetry: The congestion control algorithm will be supported by end-to-end telemetry. In other words, the cause of the congestion can be detected and shared. Action can be taken rapidly by reducing the signal path of congestion signaling to the endpoints. 

Besides these technological advances, the UEC emphasized the upcoming ethernet's scale, reliability, and security. 

“The UEC transport incorporates network security by design and can encrypt and authenticate all network traffic sent between computation endpoints in an AI training or inference job. The UEC will develop a transport protocol that leverages the proven core techniques for efficient session management, authentication, and confidentiality from modern encryption methods like IPSec and PSP,” the paper states. 

Why Improve Ethernet?

Developments in ethernet technology are nothing new. While improvements have primarily focused on increasing throughput so that more data can be transported, UEC's proposed advancements are much more profound. There’s a glaring need for improvements in hardware technologies to support the massive AI workloads. 

At PivIT, we have repeatedly discussed the challenges posed by existing infrastructure across enterprises for handling workloads. The current infrastructure is simply not built for the level of computing required for operations such as developing deep learning and large language models (LLMs). 

Multiple surveys and reports show that the tech industry isn’t fully prepared, in terms of hardware, to take full advantage of AI. 

With ethernet often used to connect processors and GPUs in computers and servers, it enables better performance for AI workloads. It only makes sense to enhance its capabilities and make it suitable for broader AI applications, not just for tech companies creating AI models but also for those using them. 

The tech companies behind the consortium have come together to address the critical need for making commonly used network components more AI-ready. 

A Rapidly Changing Hardware Landscape

Ethernet isn’t the only technology getting an AI-influenced overhaul. Servers, storage, routers, switches, and firewalls are all under the microscope for performance enhancements. Manufacturing leaders in the tech industry are ushering in an era of AI hardware that, in one way or another, facilitates the development and use of revolutionary technology. 

Many changes are happening, with vendors spending millions on redesigns and innovations to make equipment more AI-friendly. The increasing demand for better equipment with the capacity for AI is also driving this change. 

Major networking equipment makers like Cisco, Juniper, Arista, and Palo Alto are developing and releasing devices with far more advanced capabilities than they had in just a few years. It’s no less than a marvel of technological innovation. 

Of course, the changing hardware landscape will also be costly, as enterprises may need to refresh and restructure their physical infrastructure. However, the potential return on this investment can be multifold. 

PivIT is closely monitoring the developments in the hardware world. We understand the challenges enterprises face when it comes to refreshing equipment. And when developments are underway, buying new equipment can become even more confusing (not to mention very expensive).

PivIT offers a full range of hardware options with pricing to fit any budget. We also have no OEM exclusivity contracts, meaning we can get creative in developing a strategy to source the equipment you need to scale your operations.