A handful of months after showcasing the NVLink Fusion backplane at Computex, Nvidia has published its most comprehensive technical deep-dive on the tech yet, revealing exactly how the technology works under the hood.

Technical details were published during the Hot Interconnects event, outlining how NVLink Fusion can integrate with most any NIC, data processing unit (DPU), or scale-out switch to give customers building high-end AI systems the flexibility to build what they need.

The current fifth-generation NVLink supports 72-GPU topologies delivering 1,800 Gbps bandwidth and 130 Tbps of aggregate bandwidth.

The offering provides access to NVLink scale-up technologies capable of working in tandem with custom silicon, meaning hyperscalers building their own chips, like Google with its TPUs or Meta’s MTIA hardware, would be able to tap Nvidia’s scale-up fabric.

Custom hardware can be integrated via universal chiplet interconnect express (UCIe) interface, with Nvidia providing the bridge chiplet for seamless connection to NVLink fabric.

NVLink Fusion users are also able to bring together CPU offerings from the likes of Fujitsu. Custom CPU configurations can be integrated via Nvidia’s NVLink-C2C IP, with users also able to tap into CUDA-X libraries to further improve computing performance. Qualcomm is already planning to tap the technology, using it alongside its CPUs in deployments with Saudi Arabia's new Humain AI venture.

The interconnect offering is built on the Open Compute Project's (OCP) MGX rack architecture standards, providing an open foundation for custom AI infrastructure deployment.

NVLink Fusion was unveiled at Computex back in May, with the likes of MediaTek, Marvell, and Synopsys among the early adopters.

Alongside the ability to integrate a variety of hardware to power AI workloads, it also provides access to Nvidia’s ConnectX-8 SuperNICs and the Ethernet-based networking fabric Spectrum-X.

Co-packaged optics versions of both Spectrum-X and the Infiniband-based Quantum-X fabric will be added to NVLink fusion “soon,” the vendor confirmed.

“By leveraging decade-long expertise in NVLink scale-up technologies, coupled with the open, production-deployed standards of the OCP MGX rack architecture and ecosystem, NVLink Fusion empowers hyperscalers with unparalleled performance and comprehensive customization options," Nvidia noted in a Hot Interconnects blog post.

SDxCentral recently sat down with the Nvidia architects behind the Spectrum-X component of NVLink Fusion to learn about how the tech has become the foundational plumbing for some of the world's most demanding AI workloads, including xAI's Colossus supercomputing cluster.