Overview
- AMD introduced the Instinct MI455X GPU and a validated Helios rack at its Advancing AI event Thursday, saying each MI455X has 432 GB of HBM4 memory, a 2.4 GHz peak clock and that Helios can house 72 GPUs for roughly 31 TB of shared HBM4 per rack.
- The company reported large customer commitments from Microsoft, OpenAI, Meta, Oracle and Anthropic, with reported multi‑gigawatt deals that include Anthropic agreeing to deploy up to 2 GW and AMD tying up to $5 billion in milestone‑linked investment to that partnership.
- AMD detailed CDNA5 architecture changes that target higher matrix and token throughput, including chiplet XCDs, enlarged shared L2 cache, multicast/broadcast memory mechanisms, a coherent Infinity Fabric CPU link, and UALoE rack networking for single‑hop all‑to‑all communication.
- Key near‑term tests remain supply, delivery timing and software maturity because customers must validate real‑world throughput, token cost and ROCm ecosystem support before large deployments displace entrenched CUDA‑based Nvidia systems.
- If AMD converts reported orders into delivered racks and proves cost‑per‑token and software portability, Helios could loosen Nvidia’s grip on frontier AI infrastructure while raising data‑centre demand for power, space and networking capacity.