AMD + Celestica launch Helios rack platform
AMD and Celestica announced a rack‑scale 'Helios' AI platform aimed at the next era of AI compute, positioning Helios as a competitor to rack solutions like NVL72. The collaboration is AMD’s push into validated rack systems via partners. (amd.com)
Celestica will design and manufacture the high‑speed networking switches that tie Helios trays together and will act as the platform’s manufacturing partner. ((marketwatch.com)) The Helios configuration shown at CES pairs AMD Instinct MI455X accelerators with next‑generation EPYC “Venice” (Zen‑6) CPUs in tray/rack layouts built around 72‑GPU scale‑up nodes. ((tomshardware.com)) AMD’s public figures for the Helios architecture include roughly 2.9–3.0 exaFLOPS of FP4 compute per rack, approximately 31 TB of HBM4 GPU memory across the rack, and multi‑TB/s scale‑out bandwidth targets. ((phoronix.com)) AMD has said the MI400 family and Helios rack designs are on track for second‑half 2026 availability, and press coverage projects the validated, manufactured Helios platform to begin deliveries in late‑2026. ((nextplatform.com)) Helios is being developed around Open Rack Wide v3 / double‑wide form factors and UALoE scale‑up connectivity—an explicit emphasis on OCP‑style openness and switch‑based interconnects rather than a proprietary NVLink fabric. ((nextplatform.com)) By contrast, NVIDIA’s Vera Rubin NVL72 unifies 72 Blackwell GPUs with 36 Grace CPUs in a liquid‑cooled, NVLink‑centric rack offering that NVIDIA publishes as shipping today. ((nvidia.com))
