NVIDIA links custom XPUs to NVLink infrastructure

AI factories are measured by delivered output, including tokens per second, tokens per watt, cost per token, utilization and uptime. NVIDIA frames custom XPU deployment as a full-platform challenge rather than a chip-design problem, requiring scale-up and scale-out networking, rack architecture, production software and a mature supplier ecosystem.

NVLink Fusion connects XPUs to NVIDIA’s NVLink infrastructure to improve performance and reduce deployment risk. Sixth-generation NVLink supports a 72-XPU domain, with XPU-to-XPU transfers described as having 3x lower end-to-end latency than off-the-shelf Ethernet alternatives and a 10x higher packet rate. NVIDIA also points to future NVLink roadmap configurations with domains of up to 1,152 accelerators and co-packaged optics.

The platform includes NVLink-C2C for connecting XPUs to NVIDIA Vera CPUs or other ecosystem CPUs, delivering up to 6x the energy efficiency of a PCIe interface. NVLink Fusion adopters can use NVIDIA MGX rack-scale architecture, existing manufacturing partners and suppliers for rack, cooling, power and 800 VDC designs, while sharing footprints, networking, cooling, power delivery and management systems with GPU-based systems such as Vera Rubin NVL72.

NVIDIA’s DSX reference architecture and Omniverse DSX AI Factory Blueprint support facility modeling before deployment. At the rack level, reference compute trays use 100% liquid cooling and are designed for service while the rest of the rack remains operational, with NVIDIA software tools handling distributed workloads, disaggregation, cluster management, telemetry and debugging.