Redefining Artificial Intelligence Connectivity: Separating Facts from Fallacies
In a significant move for the tech industry, the Scale-Up Ethernet (SUE) framework has been introduced to the Open Compute Project, aiming to unify the industry around a single AI networking standard. This shift is driven by the recognition that if assumptions are grounded in five-year-old architectures, it's time to update the playbook.
The case for Ethernet as the preferred choice over alternatives like InfiniBand is strong. Ethernet now rivals and often outperforms InfiniBand, offering a robust ecosystem, diverse vendor options, and faster innovation. The future of AI networking, it seems, is Ethernet, and this future is already here.
One of the key advantages of Ethernet is its ability to enable flatter, more efficient topologies. As networking continues to be a core driver of AI performance, efficiency, and growth, this simplicity is a significant advantage.
Another compelling aspect of Ethernet is its flexibility and openness. Proprietary interconnects and exotic optics are no longer essential in modern AI, as Ethernet provides a broad set of choices for performance, power, and economics.
The manufacturer leading the charge in building the world's largest GPU clusters using Ethernet for scalable networks is Lambda, a leading AI infrastructure company specialising in powerful GPU clusters for advanced machine learning workloads at scale. Lambda's expansion of its GPU cloud offerings and data center capacities is supported by Nvidia’s involvement.
InfiniBand, designed for a different era, isn't equipped to handle the extreme scale seen today. In contrast, Ethernet is thriving, with 51.2T switches in production, and Broadcom's new 102.4T Tomahawk 6 setting the pace. The largest hyperscaler deployments worldwide are built on Ethernet.
Moreover, Ethernet allows for the unification of scale-up and scale-out on the same fabric, simplifying operations and enabling interface fungibility. Modern Ethernet switches, including Tomahawk 5 and 6, embed advanced load balancing, telemetry, and resiliency, reducing cost and power draw.
AI clusters can scale independently of GPU/XPU choice, ensuring openness and long-term scalability. Additionally, Ethernet supports workload-specific tuning, further enhancing its versatility.
In conclusion, Ethernet is now the standard for AI at scale, used in nearly all of the world's largest GPU clusters built in the past year for scale-out networking. The strategy of simplifying rather than over-engineering, as seen in the integration of NIC functions into XPUs themselves, reinforces Ethernet's position as the preferred choice for AI networking.
