AMD reveals industry’s first 400 Gbps Ultra Ethernet network card for AI and HPC

Cutting corners: As AI models continue to grow in size and complexity, the need for tailored networking solutions is becoming increasingly critical. AMD’s introduction of the Pollara 400 underscores the rising importance of specialized hardware in the AI ecosystem.

AMD has unveiled the Pensando Pollara 400, a fully programmable 400 Gigabit per second (Gbps) RDMA Ethernet-ready network interface card (NIC) designed to support AI cluster networking.

The rise of generative AI and LLMs has exposed critical shortcomings in conventional Ethernet networks. These advanced AI models require intense communication capabilities, including tightly coupled parallel processing, rapid data transfers, and low-latency communication. Traditional Ethernet, originally designed for general-purpose computing, has struggled to meet these specialized needs.

And yet Ethernet remains the preferred choice for AI cluster networking due to its widespread adoption. However, the growing gap between Ethernet’s capabilities and the demands of AI workloads has become increasingly evident.

AMD says its Pensando Pollara 400 is specifically designed to optimize data transfer within back-end AI networks, with a particular focus on GPU-to-GPU communication. According to AMD, the Pollara 400 can deliver up to a sixfold performance boost for AI workloads compared to traditional Ethernet solutions.

The Pollara 400 is engineered to address the communication patterns of AI workloads, offering high throughput across all available links, reduced tail latency, scalable performance, and faster job completion times.

For example, the card uses intelligent multipathing to dynamically distribute data packets across optimal routes, preventing network congestion. Its programmable hardware pipeline enables customization and optimization of network processes, while its programmable RDMA transport enhances remote direct memory access capabilities. Additionally, the Pollara 400 accelerates communication libraries commonly used in AI workloads.

AMD is releasing the Pollara 400 even though the Ultra Ethernet initiative has delayed the release of the version 1.0 specification from the third quarter of this year to the first quarter of 2025.

BOOM! AMD enters the back end AI network game with a UEC (UltraEthernet) NIC. Did not expect this one, but understand the Nvidia total solutions posture has to drive this. Will have to do some work to get underneath this. CC @WillTownTech @DellTech CTO on stage discussing. $AMD… pic.twitter.com/LBwGt34zWz

– Patrick Moorhead (@PatrickMoorhead) October 10, 2024

This new standard aims to scale Ethernet technology to meet the performance and feature requirements of AI and HPC workloads. It is designed to retain as much of the original Ethernet technology as possible while introducing profiles tailored to the specific needs of AI and HPC, which, though related, have distinct requirements.

AMD’s NIC is scheduled to begin sampling in the fourth quarter of 2024 and will be commercially available in the first half of 2025, around the time the UEC 1.0 specification is released.