HPE Slingshot vs InfiniBand vs Spectrum-X: The AI Interconnect War in Plain English
An AI cluster is a group of GPUs having a very fast conversation. Buy the best GPUs and a mediocre network, and they spend much of their time waiting for each other. The network that links servers and racks together is called the interconnect, and there is a quiet war over who supplies it.
Three names come up when you configure HPE systems: InfiniBand, NVIDIA Spectrum-X and HPE Slingshot. Each vendor says theirs is best. This guide explains what they are and what the claims are worth.
First, two different jobs
| Job | What it connects | Typical technology |
|---|---|---|
| ⬆️ Scale-up | GPUs inside one server or rack, so they act as one | NVLink (NVIDIA), UALink (open) |
| ↔️ Scale-out | Servers and racks to each other | InfiniBand, Spectrum-X, Slingshot |
This article is about the second job. For the first, see AMD Helios vs NVIDIA NVL72.
The three contenders
| Compare | InfiniBand | Spectrum-X | HPE Slingshot |
|---|---|---|---|
| 🏭 Who makes it | NVIDIA | NVIDIA | HPE from Cray |
| 🧬 What it is | A separate network type, not Ethernet | Ethernet tuned for AI | Ethernet tuned for supercomputing |
| 🔌 Works with standard Ethernet gear | No Needs gateways | Yes | Yes |
| 🏠 Natural home | AI training clusters, traditional HPC | AI clouds and enterprise AI factories | HPE Cray supercomputers |
| 🤝 Supplier independence | Low One supplier | Medium One supplier, standard protocol | Medium One supplier, standard protocol |
| 👥 Skills on the market | Specialist | Common Ethernet skills transfer | Specialist HPC |
InfiniBand: the incumbent
InfiniBand has been the default network for supercomputers and large AI training clusters for years. It was designed from the start for low latency and for not dropping data under load. NVIDIA owns the technology through its Mellanox acquisition, and pairs it with its GPUs as a complete package.
The honest criticism: it is a parallel world. It needs its own switches, its own cabling plan and its own skills, and it comes from the same supplier as your GPUs.
Spectrum-X: NVIDIA's answer to "just use Ethernet"
Many data centres would rather run everything on Ethernet, which their teams already understand. Spectrum-X is NVIDIA's Ethernet platform built for AI traffic. NVL72-class systems are offered with either NVIDIA's InfiniBand or Spectrum-X Ethernet, at up to 800 Gb/s per GPU on the GB300 generation.
The honest criticism: it speaks a standard protocol, but the performance features depend on NVIDIA switches and NVIDIA network cards working together. Standard on paper, single-vendor in practice.
Source: Supermicro GB300 NVL72 specifications
HPE Slingshot: the supercomputer's Ethernet
Slingshot is HPE's interconnect, inherited from Cray. HPE describes it as combining the specialised behaviour of HPC fabrics with the ubiquity of Ethernet. The current Slingshot 400 generation delivers 400 Gb/s per port on 64-port switches, includes automated congestion management, and is the fabric in HPE's Cray Supercomputing GX5000 platform.
HPE's marketing claim is bold: unified HPC and AI clusters of more than 260,000 endpoints, with better application performance than InfiniBand at Ethernet economics.
The honest criticism: that comparison is HPE's own. Slingshot's track record is in very large HPE Cray systems. Outside that world it is rarely an option, and the pool of engineers who know it is small.
Sources: HPE Slingshot switches, SDxCentral on Slingshot 400
The controversial part: nobody is neutral
Every benchmark in this field is published by someone selling one of the options. NVIDIA sells two of the three. HPE sells the third and resells the other two. Independent, like-for-like comparisons at scale are rare, because few organisations can afford to build the same cluster three times.
So treat "better than" claims as starting points for questions, not conclusions. The questions that matter are practical ones.
How to choose without a benchmark
| If this describes you | Start by looking at |
|---|---|
| 🏢 Enterprise team, Ethernet skills, NVIDIA racks | Spectrum-X |
| 🏋️ Large-scale training, HPC-experienced staff | InfiniBand |
| 🔬 Buying an HPE Cray supercomputer or mixed HPC and AI system | Slingshot |
| 🧩 One or two 8-GPU servers | Good standard Ethernet. You do not need a fabric war. |
| 🔓 Supplier diversity is a formal requirement | Ask every vendor about Ultra Ethernet support and roadmap |
That fourth row is worth repeating. Most organisations reading about interconnects do not have a cluster large enough for the choice to matter. Below a few dozen GPUs, the network is important but ordinary. The basics are covered in Why Your AI Project Might Need More Networking Power Than GPU Power.




