Helios racks and Vera Rubin talk
AMD and Celestica announced the Helios rack‑scale AI platform aimed at blending high‑performance compute and networking, and Nvidia’s GTC showcased the Vera Rubin Ultra platform and rack‑scale AI themes — both moves point to vendor roadmaps that could be repurposed for heavy analytics and market‑data processing announced and highlighted.
Celestica said) it will design and manufacture the “scale‑up” networking switches for AMD’s Helios architecture using the Open Compute Project Open‑Rack‑Wide (ORW) form factor to connect AMD Instinct MI450 Series GPUs. The Helios collaboration was formally announced on March 16, 2026, and Celestica confirmed) that it will lead R&D and volume manufacturing at launch to target large‑scale AI cluster deployments. NVIDIA described) Vera Rubin as a multi‑rack platform in full production that ships NVL72 GPU racks, Vera CPU racks, Groq 3 LPX inference racks, and BlueField‑4 STX storage racks as part of an integrated data‑center stack. Industry reporting attributes Rubin GPU silicon specs of roughly 336 billion transistors with 288 GB of HBM4 and claims up to ~50 petaFLOPS of FP4 inference per GPU reported). Helios emphasizes scale‑up switch silicon and ORW physical integration for AMD MI450 networking performance noted), while Vera Rubin partner systems demonstrate rack interconnects using NVIDIA Spectrum‑6 SPX Ethernet and InfiniBand‑class fabrics in validated OEM racks shown). HPE announced) Vera Rubin‑based systems and expanded private‑cloud AI offerings at GTC 2026, and NVIDIA executives projected potential platform demand on the order of $1 trillion in orders through 2027 during the keynote reported).
