Nvidia’s NVLink Fusion Bridges AI Chips and Custom CPUs

Silicon Synergy Reimagined

At the Computex 2024 keynote, Nvidia introduced NVLink Fusion—its latest high-speed interconnect designed to unify Nvidia GPUs with custom-built CPUs or third-party AI accelerators. This marks a pivot in Nvidia’s strategy to support broader system architectures beyond its own Grace CPU line. Coalescing various silicon components into a cohesive AI workload platform, NVLink Fusion is built to support exascale computing initiatives and hyperscale data centers that demand unmatched performance, low latency, and massively parallel operations. The move signals Nvidia’s acknowledgment of evolving customer needs, where data centers are increasingly mixing and matching processing elements tailored for specific AI tasks.

The Backbone of Next-Gen AI Supercomputers

NVLink Fusion offers up to 1.8TB/s bandwidth and enables coherent memory sharing between attached devices, dramatically improving efficiency for large-scale AI training and inference. Nvidia plans to extend its support to customers who want to attach homegrown silicon—such as custom Arm-based CPUs or ML accelerators—alongside its dominant GPUs, creating hybrid systems tailored to specific AI workloads. This approach mirrors the modular, flexible architecture adopted by competitors like AMD and Intel, showcasing Nvidia’s readiness to adapt while maintaining its GPU leadership. With this technology, Nvidia positions itself as a central nervous system for heterogeneous AI infrastructure.