NVIDIA Vera Rubin Supercomputing Platform Reaches Full Production to Power Global Agentic AI Factories through Advanced Infrastructure and Manufacturing Ecosystems
NVIDIA's Vera Rubin supercomputing platform has now reached full production for the rapid growth of agentic AI factories around the globe. According to an official NVIDIA announcement Today, "leaders from global supply chains and server manufacturers are building these systems at scale and are now stocking cloud providers, research laboratories, and hyperscale data centers with Vera Rubin and are powering autonomous, multi step reasoning that marks the next generation of enterprise software."
The Vera Rubin platform is a large scale out infrastructure product which has been designed to function as a single integrated supercomputer. A standard deployment of 5 racks is configured to act as a single integrated supercomputer which relies on a variety of proprietary technologies the Vera Rubin NVL72 cabinets, Vera CPUs, Groq 3 LPX processing units, Vera BlueField 4 STX storage, and Spectrum 6 SPX Ethernet networking. This architecture has the capability to deliver 10x the agent throughput at scale over its previous Grace Blackwell design.
As NVIDIA founder and chief executive Jensen Huang explained in the press release regarding this development, "Agentic workloads require a totally different infrastructure design a single prompt can spur a 1,000 step journey of reasoning, tool use, data retrieval, and response generation," he said. "The Vera Rubin engine was designed to meet the extreme computational density and power efficiency required to run these autonomous processes with immense scale, security, and predictability."
This release marks the 3 rd generation of NVIDIA’s open source MGX rack architecture. A total supply chain involved in manufacturing Vera Rubin utilizes more than 350 factories, located across 30 countries with over 150 manufacturers located just in Taiwan. Dell Technologies, HPE, Lenovo and Supermicro are all included in the roster of the industry's top enterprise infrastructure builders which will provide the platforms to end users after they achieve full production of the Vera Rubin system. Other system integrators and storage providers like ASUS, Foxconn, GIGABYTE, Hitachi Vantara, IBM, and NetApp are also in the process of creating customized systems and preparing them for market with delivery scheduled to begin in the Fall of this year.
To aid in inter cluster communications, this new system utilizes Spectrum X Ethernet Photonics which has the first co packaged optical switches powered by 200gigabit/sec SerDes. This technology results in systems that are 5 times more power efficient than other network transceivers on the market today and 5 times more uptime and is 1.3x faster to set up. By utilizing these advanced optical technologies, users will be able to easily set up systems with up to 1,000,000 GPUs, and include some major cloud provider customers such as Lambda, Oracle Cloud Infrastructure and CoreWeave who are utilizing these systems and technologies.
The system is secured using BlueField 4 data processing units with software defined networking up to 800gigabits/sec that can also support multi tenant isolation for control and security purposes while including full stack Confidential Computing capabilities to secure a fully hardware encrypted trusted execution environment all the way across the NVLink network. This can provide a strong basis for security on the platform, so major cloud providers like Microsoft Azure, IBM Cloud, Nebius and Vultr are considering integrating these technologies for securing proprietary data during active inference on models.
