To effectively train and deploy generative AI, large language models, or agentic AI, it's crucial to build parallel computing infrastructure that offers the best performance to meet the demands of AI/ML workloads but also offers the kind of flexibility that the future of AI demands. A key aspect for consideration is the ability to scale-out the intra-node GPU-GPU communication network in the data center.

At AMD, we believe in preserving customer choice by providing customers with easily scalable solutions that work across an open ecosystem, reducing total cost of ownership—without sacrificing performance. Remaining true to that ethos, last October, we announced the upcoming release of the new AMD Pensando Pollara 400 AI NIC. Today we're excited to share the industry's first fully programmable AI NIC designed with developing Ultra Ethernet Consortium (UEC) standards and features is available for purchase now. So, how has the Pensando Pollara 400 AI NIC been uniquely designed to accelerate AI workloads at scale?Building High-Performing AI Infrastructure
Cloud service providers, hyperscalers, and enterprises are looking to maximize the performance from their AI clusters. However, the network has been cited by many as a primary bottleneck for GPU utilization issues. Data transfer speed only matters if the network is properly optimized to capitalize on it.

As AI workloads continue to grow at an astounding rate, organizations cannot afford to underutilize networking and compute resources. The three main attributes of networks with the highest rates of utilization are delivering on intelligent load balancing, congestion management, fast failover and loss recovery. Highly performant networks must also be continually optimized for increased uptime, job completion times, reliability, availability and serviceability—at scale.

Extensible Future-Ready Infrastructure
Hardware Programmability Powering Customer Roadmaps
Powered by our P4 architecture, the Pensando Pollara 400 AI NIC offers a fully programmable hardware pipeline. This powers the ultimate customer flexibility-from adding new capabilities, like those issued by the UEC and/or developing custom transport protocols, designed to accelerate company roadmaps. Now as new standards, company initiatives, or AI workloads emerge, customers don't have to wait for the next generation AI NIC hardware to accelerate their workloads.

Developing UEC Features for Accelerating Next-Gen Workloads

• Transport Protocol of Choice: tap into RoCEv2, UEC RDMA, or any Ethernet protocol of your choice.
• Intelligent Packet Spray: a feature to increase network bandwidth utilization with advanced adaptive packet spraying, which is critical for managing the high bandwidth and low latency required by large AI models.
• Out-of-Order Packet Handling and In-order Message Delivery: designed to reduce buffer time by intelligently managing out-of-order packet arrivals, a common challenge associated with multipathing and packet spraying techniques, this feature minimizes errors and enhances efficiency during AI training and inference, all without relying on scale-out switching fabric.
• Selective Retransmission: enhance network performance by only resending lost or corrupted packets through in-order message delivery and selective acknowledgment (SACK) retransmission.
• Path-Aware Congestion Control: optimize network performance with intelligent load balancing automatically avoiding congested paths and help sustain near wire-rate performance during transient congestion.
• Rapid Fault Detection: accelerate AI job completion times by issues in milliseconds, with sender-based ACK monitoring, receiver-based packet monitoring, and probe-based verification, enabling near-instantaneous failover, minimizing GPU idle time.

The Open Ecosystem Advantage
By offering vendor agnostic compatibility, organizations can build AI infrastructure that meets the demands for current workloads and offers easy scalability and programmability for future requirements. With this open ecosystem approach, the AMD AI NIC helps reduce CapEx without sacrificing performance and without a dependency to deploy expensive cell based, large buffer switching fabrics.

Validated in Some of The Largest Scale-out Data Centers
Last, but certainly not least—the Pensando Pollara 400 AI NIC is set to power some of the largest scale-out infrastructure with the first customer shipments having gone to and been tested by some of the largest Cloud Service Providers (CSPs) in the world. CSPs chose the Pensando Pollara 400 AI NIC due to its distinctive programmability, high-bandwidth, low latency performance, rich feature set, and truly extensible infrastructure compatible across an open ecosystem.