# AMD said its Pollara 400 AI NIC became available for purchase AMD said on April 9, 2025 that its Pensando Pollara 400 AI NIC was available for purchase, moving the 400G programmable Ethernet component past sampling. Canonical: https://rangoon.ai/insights/archive-amd-pollara-400-availability-april-2025/ Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/) Published: 2026-10-24 Updated: 2026-10-24 Event date: 2025-04-09 Tags: [AMD](https://rangoon.ai/insights/tags/amd/), [AI networking](https://rangoon.ai/insights/tags/ai-networking/), [AI hardware](https://rangoon.ai/insights/tags/ai-hardware/) Source note: Historical reporting: AMD's dated April 9 post says Pollara 400 was available for purchase. The public product brief was freshly read as the component record. AMD performance comparisons remain vendor-attributed and do not establish another cluster's results. Image: AI NIC fabric validation. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied. ## Key takeaways - AMD's April record changed Pollara 400 from an earlier planned component to a stated purchasable AI NIC. - The company described a 400G, P4-programmable Ethernet path with RoCE and developing UEC-oriented features. - NIC availability does not prove congestion behavior, collective performance or compatibility in a target fabric. ## AMD reported a move from planned release to purchase availability [source 1](https://www.amd.com/en/blogs/2025/powering-next-gen-ai-infrastructure--a-new-era-of.html) [source 2](https://www.amd.com/content/dam/amd/en/documents/pensando-technical-docs/product-briefs/pollara-product-brief.pdf) AMD's April 9, 2025 post says the Pensando Pollara 400 AI NIC was available for purchase. The company explicitly contrasts that with its October announcement of an upcoming release. That makes availability the narrow event here, rather than treating the earlier roadmap as shipment evidence. The announcement calls the adapter a fully programmable AI NIC designed with developing Ultra Ethernet Consortium standards and features. It does not identify a particular customer's installation, acceptance result or operating scale, so those conclusions remain outside the dated product-availability claim. ## The product material describes a 400G programmable Ethernet component [source 1](https://www.amd.com/en/blogs/2025/powering-next-gen-ai-infrastructure--a-new-era-of.html) [source 2](https://www.amd.com/content/dam/amd/en/documents/pensando-technical-docs/product-briefs/pollara-product-brief.pdf) AMD's public product brief describes Pollara 400 as a 400Gb/s AI NIC for backend GPU communication. It names RDMA over Converged Ethernet, a programmable transport and congestion-control pipeline, selective retransmission, out-of-order handling and adaptive load balancing. These are component claims about intended data-path behavior. They do not establish that every switch, cable, host driver or collective library supports the same behavior. A NIC can expose capable hardware while an installed fabric uses a different transport setting, firmware or congestion policy. ## Supplier comparison data has a named test boundary [source 1](https://www.amd.com/en/blogs/2025/powering-next-gen-ai-infrastructure--a-new-era-of.html) [source 2](https://www.amd.com/content/dam/amd/en/documents/pensando-technical-docs/product-briefs/pollara-product-brief.pdf) The product brief includes AMD performance comparisons and identifies a specific test environment for some results, including MI300X systems, a Tomahawk-4-based leaf switch and a CLOS topology. Those details are useful because they prevent a headline percentage from becoming a universal fact. They are AMD measurements under the stated configuration, not independent evidence for another cluster. An evaluation should reproduce representative collective traffic, oversubscription, fault handling and workload shapes on the target hardware. It should report host software and switch settings with the result rather than attributing an entire outcome to the NIC alone. ## AMD's April status changed Pollara 400 from a planned item to a purchasable component [source 1](https://www.amd.com/en/blogs/2025/powering-next-gen-ai-infrastructure--a-new-era-of.html) [source 2](https://www.amd.com/content/dam/amd/en/documents/pensando-technical-docs/product-briefs/pollara-product-brief.pdf) AMD's April post explicitly looks back to an October announcement of an upcoming Pollara 400 release and says the NIC was then available for purchase. That makes the new fact commercial availability, not the first description of its architecture. The product brief identifies an up-to-400-Gb/s Ethernet NIC for scale-out GPU and front-end host networking, based on the Pensando P4 architecture. It describes programmable transport and congestion control, RoCE and hardware handling for functions including adaptive load balancing and selective retransmission. AMD's April post says the path can use RoCEv2, UEC RDMA or another Ethernet protocol. It also calls packet spray a feature for high-bandwidth model traffic. Those are AMD component descriptions in the availability record. The source set does not establish a final UEC conformance claim, a customer configuration or a cluster result, so the article keeps those categories outside the availability event. ## Sources 1. [AMD: Pollara 400 AI NIC availability announcement](https://www.amd.com/en/blogs/2025/powering-next-gen-ai-infrastructure--a-new-era-of.html) 2. [AMD Pensando Pollara 400 AI NIC product brief](https://www.amd.com/content/dam/amd/en/documents/pensando-technical-docs/product-briefs/pollara-product-brief.pdf) ## Related reading - [The Ultra Ethernet Consortium released its 1.0 AI-networking specification](https://rangoon.ai/insights/archive-ultra-ethernet-1-specification-june-2025/) - [GlobalFoundries’ $16 billion U.S. plan paired expansion with packaging and research](https://rangoon.ai/insights/archive-globalfoundries-us-investment-june-2025/) - [Cisco and NVIDIA proposed a two-silicon route into enterprise AI fabrics](https://rangoon.ai/insights/archive-cisco-nvidia-enterprise-networking-partnership-february-2025/) ## More information - [Documentation index](https://rangoon.ai/llms.txt) - [AI access and policies](https://rangoon.ai/ai/) - [Source repository](https://github.com/hypler-dev/rangoon)