The Ethernet standard is poised for significant enhancements that will position it as a competitive option for hosting AI workloads, potentially challenging Nvidia’s dominant InfiniBand technology over the next couple of years. According to Gartner’s recent analysis titled “Emerging Tech: Top Trends in Networking for Generative AI,” authored by director analyst Anushree Verma, Ethernet is expected to witness a surge in adoption rates among technology providers, matching and eventually surpassing InfiniBand adoption by 2026.
Key advancements driving Ethernet’s evolution include:
- RDMA over Converged Ethernet (RoCE): Facilitating direct memory access between devices via Ethernet, thereby enhancing performance and reducing CPU overhead.
- Lossless Ethernet: Introducing advanced flow control mechanisms, improved congestion handling, enhanced hashing techniques, buffering capabilities, and advanced flow telemetry to outperform modern switch capabilities.
- The Ultra Ethernet Consortium’s (UEC) specification slated for 2024: Tailored to optimize Ethernet for AI workloads, making it a compelling choice for future deployments.
Verma anticipates that the openness of Ethernet will encourage widespread adoption of these innovations by various suppliers, fostering healthy competition and offering buyers a diverse range of options.
In contrast, InfiniBand, while currently more costly than Ethernet, is projected to witness a modest increase in adoption for generative AI workloads. However, scalability limitations and specialized management requirements may deter some network designers from fully embracing this technology.
Furthermore, the emergence of optical interconnects in AI network infrastructures is foreseen to grow significantly, with a predicted adoption rate of 25% by 2030, driven by major industry players like Intel, TSMC, and HPE. This technology is expected to enhance compute cluster scalability beyond 100Tbit/sec while consuming less power compared to traditional electrical switching methods.
Additionally, the increasing utilization of PCIe, especially when coupled with servers leveraging the CXL spec for memory sharing across the bus, is set to become prevalent in AI workloads. However, widespread adoption of these technologies is estimated to commence in the coming years, with significant uptake expected around 2026.
Verma advises users to explore early adoption opportunities to gain a competitive edge by collaborating with leading technology providers during the design phase. For those considering InfiniBand, a thorough evaluation of networking options based on performance, reliability, scalability, and cost is recommended, weighing the benefits of InfiniBand-based switches against Ultra-Ethernet-based alternatives.
