How XPUs Meet a World-Class AI Factory
nvidia ·
AI factories must be built and operated as integrated systems to deliver cost-effective token generation, requiring more than standalone accelerators. Hyperscalers and AI-native companies developing custom XPUs face significant complexity in areas like scale-up networking, rack design, cooling, and supplier management. NVIDIA's NVLink Fusion aims to alleviate this by connecting custom XPUs to NVIDIA's mature AI infrastructure, providing low-latency, high-throughput scale-up domains with faster performance than Ethernet-based alternatives. The platform also supports flexible CPU choices, includes NVLink-C2C for energy-efficient CPU connectivity, and leverages an ecosystem of partners and MGX rack-scale designs. Standardizing on a unified architecture allows operators to manage risk, plan facility buildouts ahead of silicon decisions, and deploy or reprovision capacity across XPU and GPU systems as needed. NVIDIA also offers reference architectures, digital twins, and software like NCCL and Mission Control to validate and operate these factories. Thus, NVLink Fusion lets custom XPU builders combine their own innovations with proven infrastructure to create semi-custom AI factories at scale.
nvidia