AMD Pensando Pollara 400G NIC Enters AI Data Center Deployment

AMD Pensando Pollara 400G NIC Enters AI Data Center Deployment

AMD is moving its Pensando Pollara 400G AI NIC from product introduction into real-world deployment as demand for high-bandwidth networking in AI clusters accelerates.

Following the release of Ultra Ethernet Consortium (UEC) Specification 1.0, AMD announced at its Advancing AI 2025 event that Pollara supports up to 400 Gbps of network bandwidth and implements key Ultra Ethernet capabilities designed for large-scale AI and HPC workloads.

The NIC is designed to address one of the most difficult problems in modern AI infrastructure: efficiently moving enormous volumes of data between GPUs, servers, and storage systems without turning the network into the primary bottleneck.

AMD says Oracle Cloud Infrastructure (OCI) is among the first hyperscale cloud providers to deploy Pollara alongside AMD Instinct MI350-series GPUs, with customer deliveries already underway.

🚀 AMD Advancing AI 2025 and the AI Infrastructure Stack

#

AMD’s Advancing AI 2025 strategy extends beyond individual processors and accelerators. The company is building an integrated AI infrastructure stack spanning compute, networking, and software.

Instinct MI350 Series

#

The AMD Instinct MI350 series, including the MI350X and MI355X, is based on the CDNA 4 architecture and introduces support for lower-precision data formats such as FP4 and FP6.

The MI355X features 288 GB of HBM3E memory, allowing extremely large models to be accommodated on a single accelerator and reducing the need for aggressive model partitioning across GPUs.

Fifth-Generation EPYC

#

AMD’s fifth-generation EPYC processors provide the host CPU infrastructure for systems combining EPYC CPUs with Instinct accelerators.

This CPU-GPU combination forms the compute layer, while high-speed networking becomes increasingly important as the number of GPUs per cluster grows.

Pensando Pollara 400G

#

Pollara provides the networking layer for these systems.

Its combination of 400 Gbps Ethernet connectivity, RDMA, programmable packet processing, and AI-specific congestion-management mechanisms is intended to improve GPU-to-GPU and server-to-server communication at scale.

🌐 What Is the Pensando Pollara 400G NIC?

#

The Pensando Pollara 400G NIC is an AI-focused network interface developed by AMD’s Pensando networking business.

AMD acquired Pensando in 2022 for approximately $1.9 billion, expanding its portfolio beyond CPUs and GPUs into programmable networking and data-center infrastructure.

Pollara is designed around the emerging Ultra Ethernet ecosystem and targets the networking requirements of AI and high-performance computing clusters.

Its primary role is to move data between compute nodes efficiently while reducing the amount of networking work that must be performed by host CPUs.

Ultra Ethernet vs. InfiniBand

#

AI clusters have traditionally relied heavily on InfiniBand because of its low latency, RDMA capabilities, and mature collective-communication ecosystem.

Ultra Ethernet takes a different approach: it builds AI-optimized capabilities on top of Ethernet while attempting to retain the scale, interoperability, and ecosystem advantages of Ethernet-based infrastructure.

The Ultra Ethernet architecture introduces mechanisms such as:

• Intelligent packet spraying.
• Advanced congestion control.
• Flexible packet ordering.
• Selective retransmission.
• RDMA-oriented data movement.

The objective is to make Ethernet more suitable for the highly synchronized and bandwidth-intensive communication patterns generated by distributed AI workloads.

⚙️ Pollara 400G Technical Specifications

#

The combination of these capabilities is designed specifically for environments where thousands or tens of thousands of accelerators communicate simultaneously.

🧠 AI-Specific Networking Features

#

Intelligent Packet Spraying

#

AI workloads can generate massive amounts of parallel network traffic.

Intelligent packet spraying distributes packets across available network paths instead of relying exclusively on a single path.

This can improve aggregate link utilization and reduce localized congestion.

In-Order Delivery to the GPU

#

Pollara provides mechanisms for maintaining message ordering at the GPU-facing interface.

Maintaining appropriate ordering is important for workloads where communication semantics depend on the sequence in which data is consumed.

Selective Retransmission

#

Instead of retransmitting an entire communication unit after packet loss, selective retransmission allows only the missing or affected packets to be resent.

This can reduce unnecessary network traffic and improve recovery efficiency under congestion or packet loss.

Path-Aware Congestion Avoidance

#

Large AI clusters can experience highly dynamic traffic patterns as thousands of GPUs exchange data concurrently.

Pollara’s path-aware congestion avoidance considers network-path conditions when making forwarding and congestion-management decisions.

The objective is to avoid congested paths and maintain higher effective bandwidth across the cluster.

🧩 P4 Programmability Enables Custom Networking Logic

#

One of Pollara’s more important architectural features is its P4-programmable networking engine.

P4 allows network processing behavior to be customized through software-defined packet-processing logic rather than relying entirely on fixed-function hardware behavior.

For hyperscale operators, this can provide an important degree of control.

Cloud providers and large enterprises can potentially implement customized mechanisms for:

• Congestion control.
• Packet processing.
• Traffic classification.
• Data-flow management.
• Infrastructure-specific networking policies.

This programmability is particularly relevant to hyperscalers because large AI clusters can have networking requirements that differ from conventional enterprise workloads.

🔗 RDMA Reduces CPU Involvement

#

Pollara also supports Remote Direct Memory Access (RDMA).

RDMA allows data to move directly between memory regions across networked systems while minimizing involvement from host CPUs.

In GPU-centric infrastructure, this capability can be extended to communication paths involving accelerator memory, helping reduce CPU overhead and latency.

The architecture is particularly relevant to distributed AI training, where GPUs continuously exchange:

• Model parameters.
• Gradients.
• Activations.
• Intermediate tensors.
• Collective-communication data.

As cluster size increases, inefficient communication can significantly reduce accelerator utilization. High-speed RDMA networking is therefore becoming a critical component of large-scale GPU infrastructure.

☁️ Oracle Cloud Infrastructure Deployment

#

AMD announced that Oracle Cloud Infrastructure (OCI) is among the first hyperscale cloud providers to deploy Pollara.

OCI is also adopting AMD Instinct MI350-series accelerators, creating a combined compute and networking platform based on AMD technology.

AMD says Pollara deliveries to interested customers have already begun, with broader deployment at OCI planned for the second half of 2025.

Oracle’s longer-term plan includes a large-scale Z-scale AI cluster containing 131,072 Instinct MI355X GPUs.

At this scale, networking is not simply an interconnect between a handful of servers. It becomes a fundamental component of overall system performance.

The ability to efficiently distribute traffic, handle congestion, recover from packet loss, and maintain high GPU utilization becomes increasingly important as accelerator counts grow.

📈 Why 400G AI Networking Matters

#

AI model training and inference increasingly depend on distributed accelerator clusters.

Adding more GPUs does not automatically produce proportional performance gains. Communication overhead can become a limiting factor when accelerators spend too much time waiting for data or synchronization.

The networking requirements therefore scale alongside accelerator deployments.

A simplified AI cluster communication path looks like:

GPU
 │
 ▼
Pollara 400G NIC
 │
 ▼
400G Ethernet / Ultra Ethernet Fabric
 │
 ├── GPU Node
 ├── GPU Node
 ├── GPU Node
 └── GPU Node
 │
 ▼
Distributed AI Workload

At very large scales, networking performance affects effective GPU utilization, distributed-training efficiency, collective operations, and overall cluster economics.

🔬 Ultra Ethernet’s Position in AI Infrastructure

#

Ultra Ethernet represents an important attempt to evolve Ethernet for AI-scale workloads rather than replacing Ethernet’s underlying ecosystem.

Its emergence reflects a broader shift in AI infrastructure.

Traditional enterprise Ethernet was designed around relatively general-purpose networking workloads. Modern AI clusters introduce different traffic characteristics:

• Extremely high aggregate bandwidth.
• Large numbers of synchronized endpoints.
• Heavy east-west traffic.
• Distributed collective operations.
• Sensitivity to congestion.
• High demand for low-latency communication.
• Increasing reliance on RDMA.

UEC’s architecture attempts to address these requirements through mechanisms designed specifically for AI and HPC.

Future generations are expected to push beyond 400 Gbps, with the Ultra Ethernet ecosystem targeting networking speeds of 800 Gbps and beyond.

🏗️ The Broader AMD AI Infrastructure Strategy

#

Pollara illustrates AMD’s strategy of treating AI infrastructure as more than a GPU problem.

A large-scale AI system requires several tightly integrated layers:

AI Software
     ↓
Distributed Training / Inference
     ↓
Instinct GPU Accelerators
     ↓
EPYC Host CPUs
     ↓
Pensando AI NICs
     ↓
Ultra Ethernet Fabric
     ↓
Storage / Other Compute Nodes

The networking layer becomes increasingly important as GPU clusters scale from dozens of accelerators to thousands or hundreds of thousands.

AMD’s combination of Instinct accelerators, EPYC processors, Pensando networking, and Ultra Ethernet therefore positions Pollara as part of a larger AI infrastructure strategy rather than simply another 400G Ethernet adapter.

🔮 Outlook for AI Ethernet

#

The deployment of Pollara marks an important step in the development of Ethernet-based networking for large AI clusters.

InfiniBand remains a major technology for high-performance accelerator interconnects, but Ultra Ethernet offers hyperscalers and system builders an alternative based on the enormous existing Ethernet ecosystem.

Pollara’s 400 Gbps bandwidth, RDMA support, programmable packet processing, and UEC-oriented congestion-management features demonstrate how Ethernet is evolving to address the requirements of distributed AI.

If Ultra Ethernet continues to mature and scale toward 800G and beyond, AI networking could increasingly become a competitive field alongside GPUs, CPUs, and accelerator interconnect technologies.