Meta and Oracle are upgrading their AI data centers using NVIDIA’s latest Spectrum-X Ethernet switches. These switches are designed to meet the rising demands of large-scale AI systems. The upgrade is part of an open networking framework that aims to make AI training more efficient and speed up deployment across massive compute clusters. NVIDIA’s CEO Jensen Huang describes these models as transforming data centers into “giga-scale AI factories,” with Spectrum-X acting as the “nervous system” connecting millions of GPUs for training the largest models ever built.

Next-Gen Networking for AI Powerhouses

Oracle plans to incorporate Spectrum-X Ethernet switches with its Vera Rubin architecture. This will enable the company to build large-scale AI factories that connect millions of GPUs more effectively. The goal is to help customers train and deploy AI models faster than before. NVIDIA’s Spectrum-X is designed to handle the heavy data flow needed for these massive AI projects, making the entire process smoother and more reliable.

Meta is also expanding its AI infrastructure by integrating Spectrum-X switches into its Facebook Open Switching System (FBOSS). This in-house platform manages network switches at scale. Gaya Nagarajan, Meta’s vice president of networking engineering, emphasizes that the next generation of networks must be open and efficient. This helps support ever-larger AI models and serves billions of users worldwide. Adopting Spectrum-X is a key step toward achieving these goals.

Flexible, Scalable, and Energy-Friendly Systems

NVIDIA’s MGX system is a modular design that allows partners to combine different CPUs, GPUs, storage, and networking components. This flexibility helps organizations use the same system across different hardware generations. It also enables scaling up, out, and across data centers, hosting compute and switching parts. NVLink supports scale-up connections, while Spectrum-X Ethernet handles scale-out growth.

NVIDIA is also working closely with power and cooling vendors to improve energy efficiency. They are introducing power-smoothing technology to reduce electrical spikes. This innovation can cut maximum power needs by up to 30 percent. As a result, data centers can pack in more compute power within the same space, making large AI projects more sustainable and cost-effective.

Mahesh Thiagarajan from Oracle highlighted that Spectrum-X’s efficiency will help Oracle connect millions of GPUs more effectively. This will speed up AI training and deployment for their customers. Overall, these advancements are transforming data centers into true AI factories, capable of handling the biggest models and workloads of the future.

Inspired by

• https://www.artificialintelligence-news.com/news/meta-and-oracle-choose-nvidia-spectrum-x-for-ai-data-centres/