The 5-step pipeline.
How energy turns into a deployed, financed, offtake-matched cluster, and what SLYD does at every step.
How SLYD works →Platform
SLYD Cloud
New and recovered GPU systems.
NVIDIA and AMD systems through documented manufacturer and qualified channel supply, with financing and deployment coordinated on the same platform.
Explore GPU hardware →By accelerator
Infrastructure and services
Compute, hardware, and power in one book.
Browse available GPU capacity by accelerator, configuration, region, and price, or bring supply to qualified demand.
Open marketplace →Compute
Bring supply
For buyers
Start with an indicative structure.
Tell us deal size, structure, and offtake. Any range is preliminary and subject to underwriting, diligence, and documentation.
Open Configure →For lenders
GPU market trends and deployment playbooks.
Infrastructure best practices, hardware comparisons, and industry analysis from the SLYD team.
Read the blog →Company
Networking for AI clusters
Distributed training runs at the speed of its slowest path. This page covers the four networks a cluster needs, how to choose a fabric, and how to specify one precisely enough to procure.
What is AI cluster networking?
AI cluster networking connects accelerators, servers, storage, and management systems with the bandwidth, latency, topology, congestion control, resiliency, and observability required by the workload. The correct fabric depends on cluster scale, communication pattern, accelerator platform, training or inference design, storage path, facility cabling, and operational model.
A cluster has four networks, not one
Each has a different requirement, a different failure mode, and often a different technology. Budgets that treat networking as a single line item usually turn out to have priced one of these and forgotten the rest.
| Network | Connects | Optimized for | Failure shows up as |
|---|---|---|---|
| Scale-up | Accelerators inside a node or rack | Maximum bandwidth and lowest latency between accelerators in one tightly coupled domain | Poor multi-accelerator scaling within a single node or rack |
| Scale-out | Nodes into a cluster | Collective communication throughput, predictable behavior under congestion | Training that scales badly beyond one node, or unstable step times |
| Storage | Compute to the file system | Sustained throughput during load and checkpoint phases | Idle accelerators waiting on data, and slow checkpoint writes |
| Management | Provisioning, monitoring, out-of-band access | Reachability and isolation, not bandwidth | Inability to diagnose or recover a node remotely, exactly when it is needed |
The scale-up network is usually a property of the server or rack-scale platform, not a separate purchase. The other three are designed and procured for the site.
InfiniBand or Ethernet
This is the decision that attracts the most opinion and the least evidence. It is worth answering from the workload and the operating model rather than from reputation.
Large, tightly coupled training
Purpose-built for collective communication, with mature congestion behavior and a well-established reference architecture for large training clusters. Brings its own management model, tooling, and skill requirement.
Mixed estates and existing operations
Reuses existing skills, tooling, monitoring, and vendor relationships. Modern AI-oriented Ethernet fabrics have narrowed the gap considerably, and the operational continuity is worth real money.
Who will run it
A fabric nobody on the team can debug at three in the morning is a worse fabric, whatever the benchmarks say. Operational capability belongs in the decision.
Test the real pattern
For a cluster large enough that this decision matters, it is large enough to justify testing both against the actual collective communication pattern at the intended scale.
What the topology decision actually controls
Oversubscription
The ratio between endpoint bandwidth and bandwidth between switch tiers. Tightly coupled training generally wants non-blocking or lightly oversubscribed, because collectives run at the speed of the slowest path. Inference estates usually tolerate more, and the saving is real.
Switch tiers and radix
How many switch layers the design needs, which follows from port count per switch and total endpoints. Each additional tier adds cost, latency, cabling, and failure surface.
Failure domains
What stops working when one switch or one link fails, and whether a running job survives it. Worth designing deliberately rather than discovering during an outage.
Congestion behavior
How the fabric responds when many endpoints send at once, which is the normal condition during collectives rather than an edge case. This is where fabric technologies differ most in practice.
Physical layout and distance
Cable distance limits and pathway capacity constrain where racks can be placed. This regularly forces a floor plan change, and finding out late is expensive.
Growth
Whether the topology extends to the planned size without re-cabling. Designing for the eventual scale is usually cheaper than rebuilding at it.
What goes on the bill of materials
Optics and cabling are the line items most often underestimated, in both cost and lead time. On a large fabric they can rival switch spend.
- Network adapters or SuperNICs per node, at the required port count and speed
- Leaf switches sized for endpoints per rack or row
- Spine switches sized for the oversubscription target
- Optics matched to distance, connector type, and switch compatibility
- Cables, including breakout cables where port speeds differ
- Storage network components, where separate from the compute fabric
- Management switching and out-of-band access
- Spares for every part that can fail a running job
SLYD publishes no latency, port-count, throughput, or capacity figure for any switch or adapter. Those figures require the exact model, topology, traffic conditions, firmware version, and publication date to be meaningful. For performance data, use the vendor documentation for the exact part being considered. A product being discussed here is not a statement that it is available through SLYD.
Agree acceptance tests before installation
Fabric faults are cheap to find at handover and very expensive to find during a training run. Agreeing the acceptance criteria before installation is what makes it possible to reject work that does not meet them.
- Link-level verification across every cable and port
- Error counter baselines, captured and recorded
- Collective communication benchmarks at intended job scale
- Storage throughput under a realistic job pattern
- Deliberate failure injection to confirm resiliency behavior
- A documented performance baseline for future comparison
To scope a fabric
- Node count and expected growth
- Accelerator platform and its adapter requirements
- Dominant communication pattern
- Storage throughput requirement
- Physical layout and cable distances
- Existing fabric and operational skills
- Redundancy requirement
- Target timeline
Security appears in this scope only where it affects the cluster directly: management network isolation, out-of-band access control, and segmentation of the data path. Wider enterprise security architecture is a separate conversation.
AI cluster networking questions
How many networks does a GPU cluster need?
Usually four, with different requirements each. The scale-up network connects accelerators inside a node or rack, the scale-out network connects nodes into a cluster, the storage network carries data to and from the file system, and the management network handles provisioning, monitoring, and out-of-band access. Treating them as one network is a common and expensive planning error.
Should I use InfiniBand or Ethernet for the scale-out fabric?
It depends on cluster scale, communication pattern, existing operational skills, and the wider estate. InfiniBand is purpose-built for tightly coupled collective communication and brings its own management model. Modern Ethernet fabrics designed for AI have narrowed the gap and reuse existing skills and tooling. For large tightly coupled training, test both against the actual collective pattern rather than deciding on reputation.
What is oversubscription and how much is acceptable?
Oversubscription is the ratio between the bandwidth available to endpoints and the bandwidth available between switch tiers. Tightly coupled training generally wants a non-blocking or lightly oversubscribed fabric, because collective operations run at the speed of the slowest path. Inference and mixed estates usually tolerate more. The acceptable ratio is a property of the communication pattern, not a universal number.
Why do people say the network decides training performance?
Because distributed training synchronizes. Every accelerator waits for the slowest participant in a collective operation, so a single congested path, a bad cable, or an unlucky job placement slows the whole job rather than one node. That is also why congestion control, topology, and job placement matter as much as raw link speed.
What network hardware does a cluster actually need?
Adapters or SuperNICs in each node, leaf and spine switches sized for the topology, optics or cables matched to the distance and connector type, and the management switching around it. Optics and cabling are frequently underestimated in both cost and lead time, and cable distance limits can force a physical layout change late in a project.
Does SLYD publish switch and adapter performance figures?
No. A latency, port-count, or throughput figure is only meaningful with the exact model, topology, traffic conditions, firmware version, and publication date attached, and those do not survive being summarized onto a category page. For performance data, use the vendor documentation for the exact part under consideration.
How should a fabric be validated at handover?
With acceptance tests agreed before installation rather than after. Typically link-level verification across every cable, collective communication benchmarks at the intended job scale, deliberate failure injection to confirm resiliency behaves as designed, and a documented baseline to compare against later. Faults found at handover are cheap; the same faults found during a training run are not.
What information is needed to scope a fabric?
Node count and expected growth, accelerator platform and its adapter requirements, dominant communication pattern, storage throughput requirement, physical layout and cable distances, existing fabric and operational skills, redundancy requirement, and target timeline.
Scope a cluster fabric
Share the node count, accelerator platform, communication pattern, and physical layout. Those four determine the topology, and the topology determines the bill of materials.
Page updated: August 18, 2026