Nvidia NVLink Fusion delivers 14x more bandwidth than PCIe Gen5 for AI data centers

Data Flows: Nvidia developed NVLink in 2014 as an alternative technology for connecting GPUs and other computing elements in servers and data centers. Unlike PCI Express, NVLink-connected devices use mesh-style networking communication instead of relying on a central hub. Now, Nvidia is introducing an enhanced solution tailored for the next generation of AI data centers.

At this year’s Computex, Nvidia unveiled a new generation of its NVLink interconnect technology. Called NVLink Fusion, the technology is designed to support semi-custom AI infrastructure and data center deployments. The company has already partnered with an impressive list of industry leaders working to adopt the new silicon in their own products.

Alongside the recently introduced Blackwell Ultra GPU, NVLink Fusion is a key component in Nvidia’s vision of “AI factories.” The fifth generation of NVLink is integrated into two compute-dense rack models: the GB200 NVL72 and the GB300 NVL72. Together, NVLink Fusion and Nvidia’s specialized racks can deliver a total data bandwidth of 1.8 TB/s per GPU – 14 times faster than traditional PCIe Gen5 connections.

Nvidia CEO Jensen Huang described NVLink Fusion as a powerful way to bring new technology partners into the company’s AI platform and broader hardware ecosystem. For the first time in decades, Huang said at Computex, data centers are being reimagined and re-architected to embed AI into virtually every computing platform.

Nvidia’s partners for NVLink Fusion adoption include MediaTek, Marvell, Alchip Technologies, Astera Labs, Synopsys, and Cadence. Additionally, Fujitsu and Qualcomm Technologies are using the solution to integrate their CPUs with Nvidia GPUs, enabling the development of high-performance computing systems.

NVLink Fusion can be scaled up in massive cloud provider setups, Nvidia said, putting “millions” of GPUs in control of any ASIC solution. As highlighted by Synopsys CEO Sassine Ghazi, data centers are now becoming just AI platforms that need to move massive datasets quicker and in the most efficient way.

NVLink Switch chips can deliver 130 TB/s of bandwidth in one 72-accelerator NVLink domain (NVL72), Nvidia explains, while the NVLink-C2C chip-to-chip interconnection can be used to build new integrated products.

NVLink Fusion should offer ample customization options to ASIC and silicon manufacturers. At the same time, Nvidia is offering a greatly integrated product stack with its Mission Control orchestration software component. Thanks to this unified control center, enterprise customers can automate the complex management tasks required to run a modernized AI infrastructure.