AMD is gearing up for a landmark launch in the second half of 2026 with its first rack-scale GPU cluster, the Instinct MI450X IF128. This new platform is expected to be built on a 3 nm-class TSMC process and use advanced CoWoS‑L packaging, enabling each MI450X IF128 card to feature at least 288 GB of HBM4 memory. The HBM4 memory will provide up to 18 TB/s of bandwidth, with each GPU targeting approximately 50 PetaFLOPS of FP4 compute performance and consuming between 1.6 and 2.0 kW of power. AMD’s recent strategic split in its Instinct MI400 series differentiates the MI430X for high-performance computing and the MI450X for Artificial Intelligence, revealing tailored advancements for each sector.
The Artificial Intelligence-optimized MI450X comes in two configurations: ´IF64´ for conventional single-rack deployments, and the high-density ´IF128´ for large-scale installations. The IF128 system connects 128 GPUs using an Ethernet-based Infinity Fabric network. Notably, it replaces PCIe with UALink to link each GPU directly to three integrated Pensando 800 GbE network interface cards, achieving approximately 1.8 TB/s of unidirectional bandwidth per GPU and a staggering 2,304 TB/s total bandwidth across the rack.
With this architecture, the MI450X IF128 delivers a cumulative 6,400 PetaFLOPS of FP4 compute and a total of 36.9 TB of high-bandwidth memory. In comparison, the MI450X IF64 offers about half those capabilities. AMD´s advances aim to surpass NVIDIA’s anticipated ´Vera Rubin´ VR200 NVL144, which is projected to top out at 3,600 PetaFLOPS and 936 TB/s, thus providing AMD with a significant performance advantage. This leadership is expected to hold until NVIDIA’s future VR300 ‘Ultra’ NVL576 emerges, featuring 144 GPUs with four compute dies each for extreme scaling. AMD’s forthcoming rack-scale system signals a bold challenge in the ongoing race to equip Artificial Intelligence data centers with ever-more-powerful and efficient GPU clusters.
