Ethernet vs InfiniBand for AI clusters.
The fabric decision is usually made on cluster scale and operational reality, not on a benchmark. Both options build working clusters; the difference is where the effort goes.
What actually differs
| Dimension | Ethernet | InfiniBand |
|---|---|---|
| Ecosystem | Very broad, many vendors | Narrower, specialist |
| Operations | Familiar to most network teams | Distinct tooling and skills |
| Collective performance | Strong with careful lossless design | Purpose-built for it |
| Supply | Generally broader availability | Can be constrained with GPU demand |
| Reuse | Integrates with existing infrastructure | Typically a dedicated fabric |
Where scale changes the answer
For a handful of nodes, either fabric works and the decision is mostly operational. As a training cluster grows and jobs depend on tightly synchronized collective operations across many nodes, the tail latency of the fabric starts to dominate job time, and that is where InfiniBand has traditionally been chosen.
Modern high-speed Ethernet with congestion control and careful design is used at large scale too. Both are legitimate; what is not legitimate is assuming they are interchangeable at the point of purchase.
Inference and mixed clusters
Inference fleets rarely need tightly coupled collective bandwidth between nodes. They need predictable north-south capacity to serve requests and enough bandwidth to load models. Ethernet is usually the straightforward answer there.
Practical specification checklist
- Node count today, and the count the fabric must reach
- Whether jobs span nodes or run within a single node
- Existing switching and the team's operational familiarity
- Port speed per node and total uplink capacity required
- Whether optics and cabling are in scope for the quotation
- Whether the fabric choice is fixed or open to recommendation
Describe your cluster requirement.
Compute and fabric, sourced as one requirement.