Why Ethernet is Vital to Fulfilling AI’s Potential

Ethernet outperforms all other networking technologies in terms of efficiency, as measured by the lowest job completion time (JCT) for AI training and inference. It has become the go-to choice for expansive AI clusters, offering the unique capability to link over a million AI accelerators essential for scaling AI networks. The creation of AI systems is an intricate task that requires interoperability, standardization, and technology prowess among various suppliers, spotlighting the necessity of a shared, open platform to unite these components. Ethernet uniquely delivers networking solutions that ensure peak performance and scalability with cost-effectiveness and agility, due primarily to its foundation on open standards and its extensive ecosystem support.

In contrast, InfiniBand technology is commonly deployed in high-performance computing (HPC) scenarios, where attaining low unloaded latency is the primary focus. However, from our view, when it comes to AI training and inference clusters, Ethernet outperforms InfiniBand. Specifically, Ethernet can offer several advantages, including greater bandwidth, higher radix, multi-vendor standard solutions, lower cost, power-efficient economics, improved throughput, and overall better performance. As such, for large-scale AI workloads that demand high bandwidth, reliable transport, and predictable tail latency, Ethernet is the preferred choice over InfiniBand.

In our latest white paper, Why Ethernet is Vital to Fulfilling AI’s Potential, we explore why the Broadcom portfolio provides Ethernet solutions and capabilities that assure lower costs and address network congestion challenges by optimizing network efficiency, especially for scaling and optimizing fast expanding AI workloads. Broadcom solutions, such as the switch-scheduled fabric based on Jericho 3-AI switch, endpoint-scheduled fabric based on the Tomahawk switch family, and Broadcom’s 400G Ethernet NICs, enable the AI infrastructure and fabric vital to making the network the compute platform essential to fulfilling the unique distributed computing demands of AI.

This paper provides insights into:

• The evolving Ethernet landscape
• Why Ethernet meets the AI scaling challenge
• Why Broadcom Ethernet solutions are essential to AI workload optimization
• How the Broadcom E2E Ethernet portfolio provides major cost improvements and sharp power advantages
• The Ultra Ethernet Consortium’s vital role in advancing AI training and inference

Ethernet is best suited for meeting the demands of rapidly evolving networking and scaling demands of AI clusters. Broadcom E2E networking solutions are integral to ensuring the successful implementation of the largest Ethernet installations worldwide and AI clusters are clearly no exception. The foundation for the competitive advantages of Broadcom’s portfolio are the delivery of major cost and power efficiency improvements over alternative solutions. The portfolio-wide merits of Broadcom’s E2E networking solution is underpinned by the company’s strategic commitment to keeping Ethernet open. To learn more, download Why Ethernet is Vital to Fulfilling AI’s Potential today!

In partnership with:

Download Now