Overview
- AMD publicly showcased its Helios rack-scale reference design at the OCP Global Summit, positioning it for mass deployments starting in 2026.
- Each Helios rack integrates 72 Instinct MI450 accelerators delivering 2.9 exaFLOPS of FP4 compute and roughly 31 TB of HBM.
- The platform is built on Meta’s ORW double‑wide rack and OCP standards to improve power delivery, liquid cooling, and maintainability.
- Helios combines Instinct GPUs with EPYC CPUs, Pensando DPUs, and UALoE vertical plus UEC horizontal fabrics for scale-up and scale-out networking.
- Oracle Cloud Infrastructure will be the first to offer a publicly available MI450-based AI supercluster, with deployments starting in Q3 2026 and expansion planned through 2027.