AMD + Celestica roll out Helios
AMD and Celestica announced a rack‑scale 'Helios' platform aimed at rack‑level AI infrastructure with availability targeted for late 2026 (amd.com). The move highlights alternative rack offerings entering the market as competitors pitch integrated, non‑NVIDIA rack solutions.
Celestica will lead R&D, engineering and the design‑and‑manufacture of the scale‑up networking switches that form Helios’s rack fabric (corporate.celestica.com). Those scale‑up switches are specified to implement the Ultra Accelerator Link over Ethernet (UALoE) architecture for accelerator‑level connectivity rather than a proprietary NVLink‑style link (corporate.celestica.com). AMD’s Helios reference designs call out next‑gen Instinct accelerators (MI450 family) for the rack interconnect, and earlier OEM Helios builds reference the MI455X as the top‑end accelerator for high‑density racks (corporate.celestica.com). HPE and partners have already published Helios rack figures that target 72 MI455X GPUs per double‑wide rack, roughly 31 TB of HBM4 and about 260 TB/s aggregate scale‑up bandwidth with up to ~2.9 FP4 exaFLOPS per rack in the highest‑density configuration (tensorwave.com). Helios is being delivered as an open‑standards, OCP Open‑Rack‑Wide (ORW) form‑factor architecture with liquid‑cooled trays and Ethernet scale‑up switching (Pensando/Vulcano NICs and Broadcom silicon have been named in vendor integrations) rather than a closed, vendor‑proprietary rack spec (corporate.celestica.com). The companies position Helios for cloud, enterprise and research customers and Celestica exec Steven Dorwart emphasized rapid, repeatable delivery of scale‑up networking as the commercial rationale in the collaboration announcement (corporate.celestica.com).
