At the OCP Global Summit 2025, AMD presented its open-systems approach to AI data center networking, covering the three networking tiers (scale-up, scale-out, scale-across) required for million-GPU clusters. The talk detailed the AMD Pensando Pollara 400 AI NIC — currently shipping — and its P4-programmable architecture that delivers 25% RCCL performance improvement and 50% better job completion time under packet loss versus standard RoCEv2. AMD also introduced the next-generation Pensando Vulcano 800 Gbps NIC on 3nm with 8x GPU scale-out bandwidth, coming in 2026. Multi-plane networking was highlighted as a key TCO lever, reducing switch count by 40% for 100K GPU clusters.

Key Takeaways

• AI language model parameters are growing 14x per year, with Llama 4 Behemoth reaching 2 trillion parameters, driving GPU cluster sizes toward 1 million GPUs and requiring a three-tier networking architecture: scale-up (intra-rack), scale-out (intra-DC), and scale-across (multi-site).
• AMD Pensando Pollara 400 is a shipping 400 Gbps AI NIC with P4 programmability supporting RoCEv2, Ultra Ethernet, and custom transports; it delivers 25% better RCCL performance and 50% faster job completion under 0.1% packet loss versus standard RDMA.
• Multi-plane networking splits the 400 Gbps NIC port into four 100 Gbps ports across independent leaf-spine planes, reducing switch and optics capex by ~40% for 100K GPU clusters (3,000 vs. 5,000 switches) while improving fault isolation.
• The next-generation Pensando Vulcano NIC (2026) uses 3nm, delivers 800 Gbps, and enables 3.2 Tbps scale-out bandwidth per AMD MI450 GPU via 32 UAL lanes — an 8x improvement over current MI300/350 generation.
• P4 programmability enables live in-service protocol updates without traffic disruption, which AMD argues is essential for operating million-GPU clusters where pre-deployment testing at 1K–2K GPU scale cannot validate all failure modes.
• Standard RoCEv2 suffers from ECMP hash collisions, PFC-based deadlocks/livelocks, and suboptimal go-back-N retransmission; Ultra Ethernet resolves these with intelligent load balancing, lossy congestion management, and selective acknowledgement.