NVIDIA announced at GTC Taipei that its Vera Rubin platform has moved into full production, with Taiwan’s leading server makers and global supply‑chain partners building Vera Rubin systems to supply AI labs, cloud providers and hyperscalers worldwide. The company positioned Vera Rubin as a purpose‑built engine for agentic AI workloads that launch long chains of reasoning, retrieval and tool use-workloads that demand new scale and isolation models.
Vera Rubin is a POD‑scale design composed of five purpose‑built racks that operate as a single, massive AI system. The integrated configuration links NVIDIA Vera Rubin NVL72 systems, Vera CPUs, Groq 3 LPX accelerators, Vera BlueField‑4 STX storage and Spectrum‑6 SPX Ethernet racks into one coherent platform. NVIDIA says this configuration delivers roughly 10× agent throughput at scale compared with the prior NVIDIA Grace Blackwell platform.
The rollout represents the third generation of NVIDIA’s MGX rack‑scale systems and uses an open‑source MGX design. Hundreds of supply‑chain partners are participating in the ramp-about 150 in Taiwan alone — with production spreading across more than 350 factories in 30 countries. Named system builders and partners now in full‑scale production include Dell Technologies, HPE, Lenovo, Supermicro, Foxconn, Quanta Cloud Technology, Wistron, IBM, NetApp and VAST Data.
To connect large clusters, Vera Rubin introduces Spectrum‑X Ethernet Photonics: co‑packaged‑optics (CPO) based switches with 200 Gb/s SerDes that are listed as in production. NVIDIA reports Spectrum‑X delivers roughly 5× better power efficiency, 5× longer AI uptime and 1.3× faster time to deployment compared with networks using traditional transceivers, and it positions CPO networking as the foundational fabric for scaling to million‑GPU AI factories by simplifying network design and freeing power for compute.
Early ecosystem adopters named by NVIDIA include CoreWeave, Lambda and Oracle Cloud Infrastructure. Security and multi‑tenant isolation for agentic workflows are provided by BlueField‑4 DPUs and the BlueField‑4 Advanced Secure Trusted Resource Architecture. BlueField‑4 offers software‑defined networking at up to 800 Gb/s with built‑in tenant isolation to simplify operations, tighten separation and control across million‑GPU clusters — important as agentic AI handles proprietary data, regulated content and mission‑critical models. Jensen Huang said Vera Rubin was built for this moment, calling agentic AI a new kind of workload that can launch thousand‑step journeys of reasoning and tool use.
Sources
Replies (0)
No replies in this topic yet.
