The Ultra Ethernet Consortium (UEC) has delayed the release of version 1.0 of the specification from Q3 2024 to Q1 2025, but it appears that AMD is ready to announce a true network interface card for AI data centers that is ready to be deployed in Ultra Ethernet data centers. The new unit is the AMD Pensando Pollara 400, which promises up to a six-fold performance increase for AI workloads.

The AMD Pensando Pollara 400 is a 400GbE Ultra Ethernet card based on a processor designed by the company’s Pensando unit. The network processor features a programmable hardware pipeline, programmable RDMA transport, programmable congestion control, and communications library acceleration. The NIC will be sampled in the fourth quarter and will be commercially available in the first half of 2025, shortly after the Ultra Ethernet Consortium formally publishes the UEC 1.0 specification.

The AMD Pensando Pollara 400 AI NIC is designed for optimize las redes de IA y HPC a través de varias capacidades avanzadas. Una de sus características clave es la ruta múltiple inteligente, que distribuye dinámicamente paquetes de datos a través de rutas óptimas, evitando la congestión de la grid y mejorando la eficiencia general. La NIC también incluye control de congestión según la ruta, que redirige los datos lejos de rutas temporalmente congestionadas para garantizar un flujo continuo de datos de alta velocidad.

Pollara 400 is an innovative solution that offers fast failover, meaning it can detect and prevent network failures very quickly and effectively. This ensures that communication between graphics processing units (GPUs) remains uninterrupted, which is crucial for high-performance applications. performance y disponibilidad continua. Al asegurar una comunicación fluida de GPU a GPU, Pollara 400 proporciona un rendimiento robusto, optimizando al mismo tiempo la utilización de los clústeres de inteligencia artificial (IA) y minimizando la latencia que podría afectar negativamente el prosecution de datos.

Estas características avanzadas de Pollara 400 son especialmente prometedoras, ya que tienen el potencial de mejorar significativamente la escalabilidad y la confiabilidad de la infraestructura de IA. Esto hace que la solución sea ideal para implementaciones a gran escala, donde la eficiencia y la continuidad del service son de suma importancia.

In parallel, the growth of the Ultra Ethernet Consortium reflects the interest and expansion in the development of technologies networking. With 97 members, a significant increase compared to the 55 in March 2024, the consortium is working on the UEC 1.0 specification. This specification aims to scale the technology Ethernet, which is ubiquitous, in terms of performance and features, to meet the demands of artificial intelligence (AI) and high-performance computing (HPC) workloads.

La nueva especificación UEC 1.0 se beneficiará de la reutilización de la mayor cantidad posible de technology Ethernet original. Esto no solo asegura la rentabilidad del desarrollo y la implementación de nuevas tecnologías, sino que también garantiza la interoperabilidad con las infraestructuras existentes, lo cual es crucial para una adopción fluida en el mercado.

Additionally, the UEC 1.0 specification will include distinct profiles for AI and HPC. While these types of workloads share a number of similarities, they also have significant differences in their needs and characteristics. Therefore, to maximize the efficiency of each, separate protocols will be developed that specifically fit the unique requirements of AI and HPC. This will enable organizations to get the most out of their technology infrastructures, optimizing performance and ensuring that the workloads are supported. solutions are suitable for a wide range of applications and scenarios.