AMD Unveils Helios Rackscale AI Platform With 2.9 ExaFLOPS Power and Major Tech Partners Onboard
Summary
AMD unveils its Helios Rackscale AI platform, packing 72 Instinct MI455X GPUs, 2.9 exaFLOPS of FP4 compute, and 31 TB of HBM4 memory into a single reference design, with OpenAI, Meta, Oracle, and HPE among the major partners already onboard for volume deployments expected in late 2026.
Key Points
- AMD unveils the Helios Rackscale Solution, a fully integrated AI infrastructure reference design combining 72 AMD Instinct MI455X GPUs, AMD EPYC Venice CPUs, and AMD Pensando Vulcano networking, delivering up to 2.9 exaFLOPS of FP4 compute and 31 TB of HBM4 memory for frontier AI training and inference.
- Built on open industry standards including OCP Open Rack Wide, UALink, and Ultra Ethernet Consortium, the AMD Helios platform offers industry-leading scale-up bandwidth of 260 TB/s and 43 TB/s scale-out bandwidth, with modular serviceability, liquid cooling, and hardware-enforced security features.
- AMD is rapidly expanding its Helios partner ecosystem, announcing strategic collaborations with OpenAI, Meta, Oracle, HPE, TCS, Celestica, and Nutanix, with volume deployments of Helios-based systems expected in the second half of 2026.