← Back to Latest

The Network Becomes the Computer: AI-Optimized Ethernet

As AI clusters scale to hundreds of thousands of GPUs, the network has become one of the most critical — and most challenging — components of the system. Traditional data center networking was designed for client-server traffic patterns: requests from many clients to a smaller number of servers. AI networking has a fundamentally different pattern: massive, synchronized communication between thousands of accelerators, all talking to all the others simultaneously.

The response has been a new class of AI-optimized networking — hardware-accelerated Ethernet designed specifically for the traffic patterns of AI workloads, with the goal of delivering GPU-class performance over standard Ethernet infrastructure.

Why AI networking is hard

Large AI training and inference workloads involve enormous all-reduce operations: every GPU in a cluster needs to share its computed gradients with every other GPU, typically at every training step. The total data movement is staggering — and it has to happen fast, because the entire cluster is idle while waiting for the all-reduce to complete.

The traditional solution has been InfiniBand — a specialized, high-performance interconnect that delivers the low latency and high bandwidth AI workloads need. But InfiniBand is expensive, requires specialized expertise, and locks organizations into a single vendor's ecosystem.

Ethernet is the alternative: ubiquitous, commodity-priced, and familiar. But standard Ethernet was not designed for the microsecond-level synchronization that AI workloads demand. The challenge is closing that gap — making Ethernet perform like InfiniBand for AI traffic without losing Ethernet's advantages.

NVIDIA Spectrum-X: Ethernet, AI-optimized

NVIDIA Spectrum-X is the most prominent answer to this challenge. It is an end-to-end, AI-optimized Ethernet platform combining Spectrum-X Ethernet switches, SuperNICs, and software to improve the performance and efficiency of Ethernet-based AI infrastructure. The platform delivers 1.6× better AI networking performance compared with off-the-shelf Ethernet, while providing consistent, predictable performance in multi-tenant environments.

The Spectrum-6 switch, based on the 102.4 Tbps Spectrum-6 Ethernet ASIC and paired with ConnectX-9 SuperNICs supporting up to 1,600 Gb/s per GPU, is purpose-built for Vera Rubin NVL72 AI factories. NVIDIA positions Spectrum-X as a comprehensive platform — not just a switch, but a complete networking solution that includes the software to manage AI traffic patterns.

Multiplane: scaling Ethernet to massive clusters

Spectrum-X Multiplane addresses one of the hardest problems in scaling AI networking: as clusters grow beyond today's largest configurations, traditional approaches require adding a third network tier, which adds latency, slows things down unpredictably, and drives up the cost of cabling, optics, and power.

Multiplane takes a different approach: it splits each server's network connection into several independent paths, or "planes," each running its own lightweight two-tier network. The result is a flat, simple network architecture that scales to 512,000 GPUs without the added cost and complexity of a third tier.

A dedicated hardware engine inside the NVIDIA ConnectX SuperNIC manages traffic across the planes and instantly reroutes around any failure. In an eight-plane topology, if one plane fails, the network still maintains about 90% of its total bandwidth, with hardware recovery that is 11× faster than software-based multiplane load balancing. The result: 1.6× higher AI factory output.

NVLink Fusion: custom silicon into the platform

NVIDIA NVLink Fusion brings custom silicon into NVIDIA's AI infrastructure platform, enabling hyperscalers and AI-native companies to build semi-custom AI factories with greater performance, flexibility, and speed. The platform includes sixth-generation NVIDIA NVLink and NVLink Switch purpose-built scale-up networking, as well as NVLink-C2C for energy-efficient connectivity between XPUs and CPUs.

Through the NVIDIA MGX ecosystem, adopters can use production-proven rack designs, components, manufacturing partner solutions, and open, extensible software for distributed computing, disaggregated workloads, and cluster management. By standardizing GPU- and XPU-based systems on a unified architecture, NVLink Fusion helps decouple data center buildout from silicon readiness — operators can share rack footprints, networking, cooling, power delivery, and management systems, then adjust the mix of GPUs and XPUs as supply and workload requirements evolve.

The broader trend: the network as a first-class AI component

The AI-optimized Ethernet trend reflects a broader shift in how infrastructure is designed. In the past, the network was a fabric that connected compute nodes — important, but secondary to the compute itself. In AI factories, the network is a first-class component: the performance of the entire system is bounded by the speed of communication between accelerators, and networking is codesigned with compute, not bolted on afterward.

This is visible across the industry. AMD's MI355X and Helios platforms are designed with networking in mind. AWS Trainium and Inferentia are paired with AWS's internal networking. SiFive is working with NVIDIA to port CUDA to RISC-V hardware and integrate NVLink Fusion. The networking fabric is no longer an afterthought — it is part of the architectural design from the start.

What this means for the industry

AI-optimized Ethernet is significant because it promises to democratize high-performance AI networking. InfiniBand will remain the choice for organizations that need the absolute best performance and are willing to pay for it and manage its complexity. But Spectrum-X and similar AI-optimized Ethernet platforms offer a path to near-InfiniBand performance over standard Ethernet infrastructure — which is cheaper, more familiar, and less vendor-locked.

As AI clusters continue to scale — toward 512,000 GPUs and beyond — the network will only become more critical. The organizations that figure out how to build and operate AI-optimized networking efficiently will have a significant advantage in the cost and performance of their AI infrastructure.