NVIDIA Networking
The end-to-end 800Gb/s networking platform for large-scale AI infrastructure.
Ultra-high-speed interconnects designed for trillion-parameter AI models. In AI training, thousands to tens of thousands of GPUs exchange data at every step — the network directly determines performance. Spanning InfiniBand, Spectrum-X Ethernet, DPUs, SuperNICs, and NVLink, it eliminates GPU-to-GPU communication bottlenecks.
Why AI Networking Matters
In large-scale distributed training, collective communication between GPUs (all-reduce and similar) becomes as much of a bottleneck as GPU compute itself. NVIDIA connects GPUs inside the rack with NVLink and racks to each other with InfiniBand or Spectrum-X Ethernet, and offloads communication to DPUs and SuperNICs so GPUs can focus purely on compute.
InfiniBand vs Spectrum-X Ethernet
| Aspect | InfiniBand (Quantum-X800) | Spectrum-X Ethernet (SN5600) |
|---|---|---|
| Approach | Dedicated lossless fabric | RoCEv2-based lossless Ethernet |
| In-network computing | SHARPv4 — offloads collective operations to the switch | Adaptive routing and congestion control optimized for AI traffic |
| Port speed | 800Gb/s | 800Gb/s |
| Best environment | The largest dedicated AI supercomputers and HPC | Multi-tenant generative AI clouds |
| Ecosystem | InfiniBand standard | Standard Ethernet — operational familiarity and versatility |
Networking Lineup (Specs by Line)
Quantum-X800 InfiniBand
Next-generation InfiniBand switch at 800Gb/s per port · SHARPv4 in-network computing — the fabric for the largest AI/HPC deployments
Spectrum-X Ethernet (SN5600)
800Gb/s AI-optimized Ethernet switch · adaptive routing and congestion control — for multi-tenant generative AI clouds
NVLink Switch (5th gen)
1.8TB/s per GPU · NVL72 domain — connects the GPUs inside a rack at ultra-high bandwidth as one giant GPU
BlueField-3 DPU
400Gb/s · a data processing unit that offloads, accelerates, and isolates networking, storage, and security
ConnectX-8 SuperNIC
800Gb/s · the SuperNIC that connects GPU servers to the AI fabric with GPUDirect and RoCE
LinkX
High-performance cables and optical transceivers (active/passive) for 800G infrastructure
Key Spec Comparison
| Product | Type | Speed | Key Capability | Role |
|---|---|---|---|---|
| Quantum-X800 | InfiniBand switch | 800Gb/s per port | SHARPv4 in-network computing | Largest-scale AI/HPC fabric |
| Spectrum-X (SN5600) | Ethernet switch | 800Gb/s | Adaptive routing · congestion control | AI-optimized Ethernet clouds |
| NVLink Switch | GPU interconnect | 1.8TB/s (per GPU) | NVL72 domain | Ultra-high-bandwidth in-rack GPU connectivity |
| BlueField-3 | DPU | 400Gb/s | Networking, storage & security offload | Infrastructure acceleration & isolation |
| ConnectX-8 | SuperNIC | 800Gb/s | GPUDirect · RoCE | GPU-to-fabric connectivity |
| LinkX | Cables · transceivers | 800G | Active/passive optics & copper | Physical-layer connectivity |
* A roadmap has been announced that further improves power efficiency and scalability with Quantum-X / Spectrum-X Photonics (silicon photonics · co-packaged optics).