How NVIDIA NVLink is Transforming HPC and AI: Architecture, Switches, and Network Comparisons

What Is NVIDIA NVLink

NVLink is a high‑bandwidth, low‑latency interconnect protocol that directly links GPUs within a server, overcoming the bandwidth limitations of traditional PCIe switches. The fourth generation of NVLink delivers up to 112 Gbps per lane, roughly three times faster than a PCIe Gen5 lane.

NVLink provides a simplified, point‑to‑point network for GPU‑to‑GPU communication, reducing overhead compared with conventional networks and enabling higher CUDA acceleration across multiple GPU tiers. The technology has progressed from NVLink 1.0 on the P100 to NVLink 4.0 on the H100, with each generation improving connection methods, bandwidth, and performance.

NVSwitch Chip

The NVSwitch chip acts as a high‑speed ASIC switch that aggregates multiple NVLink connections, allowing many GPUs to communicate with a combined bandwidth of up to 900 GB/s per GPU pair. The third‑generation NVSwitch (NVSwitch 3) features 64 NVLink 4 ports, delivering 12.8 Tbps of unidirectional bandwidth (3.2 TB/s bidirectional) and integrates SHARP functionality to aggregate and update computation results across GPUs, reducing network traffic.

NVLink Servers

NVLink‑enabled servers combine NVLink and NVSwitch to provide high‑performance GPU interconnects. They are found in NVIDIA’s DGX series and OEM HGX platforms. In 2022 NVIDIA announced the DGX H100, the first AI platform built on the fourth‑generation DGX system, offering unprecedented scalability for scientific computing, AI, and big‑data workloads.

These servers deliver powerful GPU interconnectivity, scalability, and the compute density required by modern HPC and AI applications.

NVLink Switch

In 2022 NVIDIA released a standalone NVLink switch built around the NVSwitch chip. The 1U device provides 32 OSFP ports, each containing eight 112 Gb/s PAM4 lanes, and houses two NVSwitch 3 chips, enabling GPU clusters across multiple hosts.

NVLink Network

By interconnecting NVLink‑enabled servers with NVSwitch hardware, a large‑scale fabric network is formed. Each server retains its own address space, providing isolated, secure GPU data paths. The network is automatically configured via software APIs at boot and can be re‑programmed during runtime.

The NVLink fabric offers higher bandwidth and lower latency than traditional Ethernet, creating a dedicated GPU‑centric network.

InfiniBand vs. NVLink Networks

InfiniBand is an open‑standard, multi‑channel, high‑speed serial network supporting point‑to‑point and multicast communication, widely used in HPC clusters. NVLink is a proprietary NVIDIA technology focused on direct GPU‑to‑GPU links.

InfiniBand excels in large‑scale data‑center deployments, while NVLink provides superior bandwidth and latency for GPU clusters in AI and HPC workloads. A benchmark comparing H100 (NVLink) with A100 (InfiniBand) illustrates the bandwidth advantage of NVLink.

Conclusion

NVIDIA NVLink has become a cornerstone technology for high‑performance computing and artificial intelligence, dramatically improving GPU communication, performance, and parallel processing capabilities. As advanced computing continues to evolve, NVLink’s role will expand, driving further innovation across HPC, AI, and data‑center domains.