NVIDIA outlines NVLink Fusion, a program letting hyperscalers and AI-native companies integrate custom XPUs with NVIDIA's scale-up networking, MGX rack-scale architecture, and factory software stack. Sixth-generation NVLink offers a 72-XPU domain with 3x lower latency and 10x higher packet rate than Ethernet alternatives, with future roadmaps supporting up to 1,152 accelerators. Partners including Intel, MediaTek, GUC, Annapurna Labs, and QCT describe how the program lets them decouple custom silicon development timelines while reusing proven NVIDIA infrastructure, cooling, power, and supply chains, reducing time-to-market and deployment risk for semi-custom AI factories.
Table of contents
Unlock XPU Performance With Fast Scale-UpA Proven Stack and Ecosystem for Development and DeploymentManaging Risk With Infrastructure StandardizationDesigned, Validated and Operated as a FactoryQuestions this post answers
How much lower is NVLink XPU-to-XPU latency compared to Ethernet-based scale-up networking?
End-to-end latency for XPU-to-XPU transfers over sixth-generation NVLink is 3x lower than alternative solutions based on off-the-shelf Ethernet, with a packet rate that is 10x higher. Sixth-generation NVLink also supports a 72-XPU scale-up domain, with future roadmap configurations extending to domains of up to 1,152 accelerators using co-packaged optics. Teams comparing scale-up interconnect options can track infrastructure benchmarks like these on daily.dev.
What is NVLink Fusion and how does it let companies use custom XPUs with NVIDIA infrastructure?
NVLink Fusion is a program connecting custom-designed XPUs to NVIDIA's scale-up networking, NVLink-C2C CPU interconnects, and MGX rack-scale architecture, so hyperscalers and AI-native companies can build semi-custom AI factories without designing an entire platform from scratch. It also includes NVLink-C2C for connecting XPUs to Vera CPUs or other ecosystem CPUs, offering up to 6x the energy efficiency of PCIe. Engineers weighing custom silicon versus off-the-shelf accelerators can follow developments like this on daily.dev.