Surgical Robotics Giants Shift Focus Away From Competition With Market Leader
The surgical robotics industry has witnessed significant growth in recent years, with companies …
14. March 2025
Revolutionizing Ai: How Ethernet Is Fueling The Next Generation Of Scalable Infrastructure
The Evolution of Ethernet in Artificial Intelligence Infrastructure: Unveiling a New Era of Scalability and Innovation
In recent years, the artificial intelligence (AI) industry has faced unprecedented challenges, driven by surging demand for AI training and inference workloads. The need for massive compute power and high-performance networking solutions has become increasingly pressing, with many organizations struggling to keep pace with the accelerating pace of innovation.
A significant shift is now underway, as organizations begin to adopt open Ethernet-based networking solutions. Unlike InfiniBand, which is controlled by a single vendor, Ethernet offers a multi-vendor ecosystem that provides greater flexibility, scalability, and resilience. This transition towards an open Ethernet ecosystem is poised to revolutionize the way AI infrastructure is built and deployed.
One of the primary benefits of Ethernet is its widespread adoption across industries. From data centers to edge computing, Ethernet has become the de facto standard for networking solutions. Additionally, its cost-effectiveness and ability to scale make it an attractive alternative to InfiniBand’s proprietary solution. As a result, organizations can now diversify their sourcing strategies, ensuring interoperability and reducing the risks associated with long lead times.
To address the performance limitations of traditional Ethernet, efforts led by the Ultra Ethernet Consortium (UEC) are defining new standards that make Ethernet AI-ready. Key innovations include NIC-based scheduling and fabric-based scheduling, which enable congestion management at the edges of the network and ensure predictable latency.
Network Interface Controller (NIC)-based solutions perform congestion management at the edges of the Ethernet network. These solutions constantly monitor network performance and take actions to minimize congestion and packet drops. By addressing this critical bottleneck, NIC-based scheduling enables organizations to maintain high-performance networking even under heavy loads.
Fabric-based scheduling is transforming Ethernet from an unpredictable, lossy technology into a predictable and lossless one. This advanced scheduling solution ensures that every piece of data entering the network is delivered to its destination with minimal latency. By achieving predictable latency, fabric-based scheduling enables all GPUs to be fully utilized, maximizing compute performance.
As AI adoption continues to accelerate, organizations are seeking scalable, cost-effective, and widely available networking solutions. The move towards an open Ethernet ecosystem addresses supply chain constraints and availability challenges that have hindered AI infrastructure in recent years. In addition to broadening the ecosystem of networking gear, Ethernet allows a flexible choice of optics and NIC vendors – often a major bottleneck.
The forecast report from 650 Group predicts that Ethernet revenue will grow from 25% of Data Center AI/HPC Backend Networking in 2024 to 61% in 2025 and up to 75% in 2026. This trend suggests that Ethernet is poised to play a dominant role in the future of AI networking.
The transition towards an open Ethernet ecosystem offers several benefits for AI infrastructure builders:
- Greater flexibility: By adopting open Ethernet, organizations can diversify their sourcing strategies and reduce reliance on single vendors.
- Improved interoperability: Multiple vendors contribute to the development of Ethernet-based networking solutions, ensuring seamless integration with existing infrastructure.
- Reduced lead times: Open Ethernet enables faster deployment timelines, as supply chains are no longer constrained by a single vendor’s limitations.
As the open Ethernet ecosystem expands, AI cloud builders and enterprises can expect greater availability, lower costs, and improved performance. This will enable them to build more scalable, resilient, and innovative AI infrastructure landscapes.
While the transition towards an open Ethernet ecosystem presents several challenges, it also offers opportunities for innovation and growth. As organizations navigate this new landscape, they must consider the following:
- Standardization: The adoption of standardized Ethernet solutions will be crucial in ensuring interoperability across multiple vendors.
- Security: As with any networking solution, security will remain a critical concern. Organizations must prioritize secure networking practices to protect against potential threats.
- Education and training: The open nature of Ethernet-based networking requires education and training for IT professionals. Organizations must invest in professional development programs to ensure that staff can effectively deploy and manage these solutions.
In conclusion, the evolution of Ethernet in AI infrastructure marks a significant shift towards a more scalable, innovative, and resilient AI landscape. By adopting open Ethernet-based networking solutions, organizations can overcome supply chain constraints and availability challenges, driving growth and competitiveness in the AI industry.