NVLink vs UALink vs PCIe: AI Interconnects Explained

NVLink, UALink and PCIe all move data between chips, but they sit at different levels. PCIe is the general-purpose bus that connects a GPU to the host. NVLink is Nvidia's proprietary link for joining GPUs into one fast memory domain. UALink is an open standard from a multi-vendor consortium that aims to do the same job for non-Nvidia accelerators. As of October 2026, NVLink is deployed in Nvidia's rack-scale systems, UALink has published specifications, and PCIe remains the baseline all of them are measured against.

Scale-up and scale-out: the idea behind the comparison

AI clusters use two kinds of network. Scale-up links tie a few dozen to a few hundred accelerators into a tightly coupled group that behaves like one big machine, with very high bandwidth and low latency. Scale-out links connect many such groups across a data center, usually over InfiniBand or Ethernet.

NVLink and UALink are scale-up technologies. PCIe is not designed for either role, though it handles the connection between GPUs, CPUs, SSDs and network cards inside a server. InfiniBand and Ethernet (including the Ultra Ethernet effort) are the scale-out side, which is why searches like "NVLink vs InfiniBand" compare tools for different layers rather than direct rivals.

PCIe: the baseline

PCI Express is an open standard maintained by PCI-SIG and found in practically every server and PC. Its x16 slot is the usual GPU connection. Figures below are per direction for a x16 link, from the PCI Express overview:

• PCIe 5.0: 32 GT/s per lane, about 63 GB/s.
• PCIe 6.0: 64 GT/s per lane, about 121 GB/s.
• PCIe 7.0: 128 GT/s per lane, about 242 GB/s.
• PCIe 8.0: 256 GT/s per lane, about 484 GB/s, planned for around 2028.

Each generation doubles the previous one, which is fast by any normal standard. For tightly coupled AI training it is still not enough, and PCIe was built around a host-and-device model rather than many peers sharing memory. Our PCIe generations guide covers the full history.

NVLink: Nvidia's GPU-to-GPU fabric

NVLink connects Nvidia GPUs directly to each other, and through NVLink Switch chips it extends that connection across a whole rack. Every GPU in the domain can reach every other GPU at full link speed. Nvidia's NVLink page lists these per-GPU figures:

• NVLink 4 (Hopper): 900 GB/s, 18 links.
• NVLink 5 (Blackwell): 1,800 GB/s, 18 links.
• NVLink 6 (Vera Rubin): 3,000 GB/s, 36 links. Nvidia marks the sixth-generation numbers as preliminary.

Nvidia's headline NVLink numbers describe total per-GPU bandwidth, so compare them with care against per-direction PCIe figures. On that page Nvidia claims sixth-generation NVLink offers 12 times the bandwidth of PCIe Gen6, a vendor comparison rather than an independent test.

The rack-scale systems built on this are the NVL72 designs, which link 72 GPUs in one NVLink domain. The switch also performs some in-network math, such as reductions, so GPUs spend less time exchanging partial results. For the GPU lineage behind each generation, see our Nvidia architectures timeline.

NVLink Fusion

NVLink is proprietary, but Nvidia now offers NVLink Fusion, a way for hyperscalers to pair semi-custom chips (custom accelerators or CPUs) with Nvidia's NVLink and rack architecture. It is a route into Nvidia's ecosystem, not an open standard: the fabric, switches and rack design remain Nvidia's. Nvidia's page does not set out licensing terms, so do not assume any specific cost or contract structure.

UALink: the open alternative

The UALink Consortium was formed in 2024 by a promoter group of AMD, Broadcom, Cisco, Google, Hewlett Packard Enterprise, Intel, Meta and Microsoft. Alibaba, Apple and Synopsys later joined its board. Nvidia is not among those names. The aim is a vendor-neutral scale-up link so that accelerators from different companies can share one fabric.

The UALink 200G 1.0 specification was published in 2025. It is based on a 200 Gb/s-class Ethernet physical layer (IEEE 802.3dj), and a single pod can contain up to 1,024 accelerators, according to the UALink article. The consortium's website announces that it published four further specifications in April 2026.

The main practical question is timing. A published specification is not a shipping product. Buyers should check which accelerators and switches actually support UALink before assuming it is available in volume; this article does not confirm any shipping UALink products.

Side-by-side summary

• Role: PCIe connects a device to a host; NVLink and UALink connect accelerators to each other.
• Openness: PCIe and UALink are open standards; NVLink is Nvidia's, with NVLink Fusion as a route into its ecosystem.
• Scale: NVLink today spans 72-GPU racks; UALink 1.0 allows up to 1,024 accelerators per pod on paper.
• Availability: PCIe and NVLink ship now; UALink is defined by published specifications; check product availability separately.
• Bandwidth: NVLink 5 is listed at 1,800 GB/s per GPU, well above a PCIe 6.0 x16 slot at about 121 GB/s per direction.

Why it matters beyond the spec sheets

Large models are split across many accelerators, and the speed of the link between them often limits training and inference. That makes the scale-up fabric a strategic choice as much as a technical one. A closed fabric ties customers to one chip vendor; an open one lets cloud providers mix suppliers. Memory and packaging matter for the same reason, as covered in our guides to HBM4 and CoWoS packaging. More coverage sits in the Chips and Big Tech sections.

Frequently asked questions

What is UALink?

UALink is an open interconnect standard from the UALink Consortium for linking AI accelerators into one scale-up network. Version 1.0 supports up to 1,024 accelerators in a pod.

Is NVLink faster than PCIe?

Yes. Nvidia lists NVLink 5 at 1,800 GB/s per GPU, while a PCIe 6.0 x16 link carries about 121 GB/s per direction. Nvidia's figure counts total bandwidth, so the gap is somewhat smaller on a per-direction basis.

Is NVLink the same as InfiniBand?

No. NVLink is a scale-up link between GPUs inside a rack-sized domain, while InfiniBand is a scale-out network that connects servers and racks across a data center.

Can other companies use NVLink?

Through NVLink Fusion, Nvidia lets hyperscalers connect semi-custom CPUs and accelerators to its NVLink ecosystem. It is not an open standard, and the technology stays under Nvidia's control.

Will UALink replace NVLink?

That is unknown. UALink is a competing open approach backed by many large companies, but NVLink is already deployed, and UALink's product availability is still to be established.