Summary:
- The Ultra Accelerator Link (UALink) consortium released the UALink 200G 1.0 Specification for AI clusters.
- Arista’s ESUN framework aims to enhance Ethernet scalability for AI collectives.
- Gartner predicts significant growth in Scale-Up AI Fabrics (SAIF) for AI infrastructure.
Article:
The Ultra Accelerator Link (UALink) consortium, comprised of major tech players like AMD, Cisco, and Microsoft, has unveiled the UALink 200G 1.0 Specification, designed to provide an open standard interconnect for AI clusters. This specification enables a data rate of 200 Giga transfers per second (GT/s) per channel, facilitating communication between accelerators and switches within AI computing pods.
Arista Networks has introduced the ESUN framework, a modular approach to Ethernet scale-up that emphasizes interoperability and performance. ESUN leverages common Ethernet headers for seamless integration of upper-layer protocols, while also incorporating an open Ethernet data link layer to support high-performance AI collectives at XPU cluster scale. By utilizing Ethernet PHY layer technology, ESUN ensures interoperability across various vendors and interconnect options.
Gartner’s research predicts a surge in the adoption of Scale-Up AI Fabrics (SAIF) to bolster AI infrastructure initiatives in the coming years. SAIF offers high-bandwidth, low-latency network interconnectivity and improved memory interaction between AI processors. While current SAIF implementations are vendor-specific and limited in proximity, Gartner recommends the use of Ethernet for connecting multiple SAIF systems to maximize scalability, performance, and supportability in AI networking fabrics.
