How Many Switches Do You Need for an AI GPU Cluster? 400G/800G Leaf-Spine Sizing and BOM

Follow Us:
Quick Take
There is no single switch count that fits every AI GPU cluster. The required number depends on GPU server count, network ports per server, 400G or 800G link design, leaf port density, spine topology, storage and management networks, redundancy, and future expansion. Cisco reference architectures provide useful examples, but their switch and BOM counts should not be treated as a universal formula.

The main mistake in AI cluster sizing is to start with the GPU count and immediately divide it by switch ports. A reliable design must first identify the number of GPU servers, the number of network links per server, the required fabric layers, and the separate storage and management networks. This guide explains how to estimate switch, optic, DAC/AOC, and breakout requirements for 400G and 800G AI Ethernet fabrics.

For a broader architecture view, see our AI GPU cluster networking solutions covering Ethernet fabrics, high-speed optics, DAC/AOC, and multi-rack deployment planning.

Start With GPU Servers, Not Only GPU Count
  • Why GPU count alone cannot determine switch count
What Networks Should Be Included in the BOM?
  • Compute, storage, management, and service networks
How Many Compute Switches Are Needed?
  • Leaf, spine, and Super-Spine sizing examples
400G or 800G: Which Should You Choose?
  • Port density, breakout, reach, and expansion considerations
How to Count Transceivers and Cables
  • Optics, DAC/AOC, breakout, and spare planning
Cisco Reference BOM Examples
  • How reference architecture counts should be interpreted
A Practical Sizing Workflow
  • Steps for turning GPU requirements into a BOM
RoCE and Procurement Checks
  • Lossless Ethernet and quotation requirements
Frequently Asked Questions
  • Common AI GPU cluster sizing questions
Final Takeaway
  • The practical rule for AI cluster switch sizing

Start With GPU Servers, Not Only GPU Count

The first sizing mistake is to start with the number of GPUs and immediately divide by switch ports. The correct starting point is the number of GPU servers and the number of high-speed network links used by each server.

Use the following basic calculation:

GPU server count = Total GPU count ÷ GPUs per server

A cluster with 1,024 GPUs could use 128 servers with 8 GPUs per server, or 256 servers with 4 GPUs per server. These designs do not create the same number of network links. The server configuration, network adapter count, port speed, and rail design determine the access-side port requirement.

A server with multiple 400G or 800G connections can consume several leaf ports even when the total server count remains unchanged.

What Networks Should Be Included in the BOM?

A complete AI cluster BOM should not only count the network used for GPU-to-GPU communication. A production cluster may include several separate or partially converged networks.

Network Main Function Planning Question
East-West compute fabric GPU collective communication and distributed training How many high-speed links connect each GPU server?
Storage network Dataset access, checkpoints, and parallel storage Is storage traffic isolated from GPU traffic?
Management network Provisioning, orchestration, monitoring, and control Which nodes and devices require management access?
Out-of-band network BMC, console, and infrastructure management Are server and switch management ports separated?
PDU and service network Power control, console access, and operational services Are power and support systems included in the BOM?

Cisco’s reference architecture separates the East-West compute network from the North-South storage and management network. It also describes separate switch management, BlueField management, and miscellaneous service networks.

This is why a BOM that only lists GPU-facing leaf switches is usually incomplete. For broader compute, storage, and network planning, see the AI cluster network, storage, and compute planning guide.

Need a preliminary AI cluster BOM?

Send your GPU count, server configuration, link speed, storage needs, and growth plan.

How Many Compute Switches Are Needed?

Cisco’s reference architecture provides an example of a three-tier East-West compute network using high-density Cisco 8000 platforms. The architecture uses Leaf switches, Spine switches, Super-Spine switches, 800G switch-to-switch links, and 400G server-to-leaf connections.

For larger AI fabrics, compare the available Cisco 8000 Series switches by port density, uplink speed, supported breakout modes, and exact hardware PID.

GPU Nodes GPUs Leaf Switches Spine Switches Super-Spine Switches
128 1,024 16 16 8
256 2,048 32 32 16
512 4,096 64 64 32
1,024 8,192 128 128 64
2,048 16,384 256 256 128
4,096 32,768 512 512 256

These figures describe the compute fabric in a specific Cisco reference architecture. They should not be copied directly into a smaller enterprise cluster. A different server NIC layout, port density, or topology can produce a different result.

400G or 800G: Which Should You Choose?

The choice between 400G and 800G depends on cluster scale, server interface design, switch radix, and expansion requirements.

Consideration 400G Design 800G Design
Typical role GPU server links, storage, and medium-to-large fabrics High-density leaf, spine, and large-scale AI fabrics
Port density Easier to deploy in many current server and switch designs Higher bandwidth per physical port
Breakout use 400G to 4 × 100G on compatible platforms 800G to 2 × 400G on compatible platforms
Procurement complexity More familiar optic and cabling options Requires strict OSFP/QSFP-DD and breakout validation
Expansion model Suitable for phased deployment Useful when high-density growth is planned from the beginning

A 400G deployment is not automatically smaller or cheaper. If every server requires multiple 400G ports, the total number of server-facing links can still be substantial.

An 800G platform may reduce the number of physical switch ports in some spine or leaf roles, but the final result depends on how the ports are broken out and how much bandwidth each server requires.

How to Count Transceivers and Cables

Every physical link should be counted from both ends. For each connection, confirm the port speed, lane configuration, optical module or cable type, fiber distance, breakout design, and FEC requirements.

Link Type Device A Device B Items to Count
GPU server to leaf GPU NIC Compute leaf One module or cable assembly per endpoint
Leaf to spine Leaf switch Spine switch Two switch-side endpoints per link
Spine to Super-Spine Spine switch Super-Spine 400G/800G links according to the topology
Storage node to storage leaf Storage NIC Storage leaf Storage-side optics or DAC/AOC
Management and BMC Server management port Management switch Management cables and port adapters

For a compatible 400G optical connection, a product such as QDD-400G-DR4-S may be considered after checking the switch PID, fiber type, reach, coding, and supported FEC.

For very short in-rack connections, a passive cable such as QDD-400-CU1M may be considered after checking port compatibility and supported cable length.

A 400G-to-400G connection, a 400G-to-4 × 100G breakout, and an 800G-to-2 × 400G breakout do not use the same BOM structure. The port configuration and every resulting endpoint must be included.

For the selection tradeoffs between copper DAC, fiber AOC, and optical modules, see the 400G DAC, AOC, and optical link guide.

Cisco Reference BOM Examples

Cisco’s published BOM table gives an example of how a complete reference architecture grows from 4K to 32K GPUs.

GPUs HGX Server Nodes 800G Switches 400G Switches Low-Speed Management Switches
4,096 512 184 12 36
8,192 1,024 378 26 72
16,384 2,048 754 50 144
32,768 4,096 1,506 98 288

These are minimum BOM values for Cisco’s defined architecture, not a generic estimate for all AI deployments. The reference design uses specific Cisco and NVIDIA hardware, server configurations, link speeds, and topology assumptions.

The table also shows why the full cluster BOM includes more than the main compute leaf and spine switches. Storage, management, and lower-speed infrastructure switches can contribute a significant number of devices.

Cisco’s current 8000 Cloud Reference Architecture should be used to verify the scope and assumptions behind these example counts.

A Practical Sizing Workflow

Define the Compute Unit

  • Record the total GPU count.
  • Confirm the number of GPUs per server.
  • Calculate the number of GPU servers.
  • Identify the NIC model and port count per server.
  • Record the target port speed.
  • Confirm the number of independent GPU network rails.

Define the Fabric

  • Choose a two-tier or three-tier topology.
  • Confirm leaf port density.
  • Confirm spine radix and uplink requirements.
  • Define the oversubscription target.
  • Document the redundancy model.
  • Include the planned growth target.

Define the Non-Compute Networks

  • Storage leaf switches
  • Storage spine or uplink requirements
  • Management switches
  • BMC and out-of-band access
  • Switch management
  • PDU and console connectivity

Define the Physical Interconnects

  • 400G and 800G optics
  • 100G or 200G breakout links
  • DAC and AOC lengths
  • Single-mode or multimode fiber
  • FEC and lane requirements
  • Spare modules and cables

Validate the Platform

  • Exact switch PID
  • Supported network operating system
  • Port breakout support
  • Optic and cable compatibility
  • Buffer and congestion-management features
  • PFC, ECN, and MTU support
  • Power and airflow requirements
  • Software and support entitlement

RoCE and Procurement Checks

If the GPU fabric uses RoCEv2, switch sizing alone is not enough. The design should also verify PFC behavior, ECN marking, queue and buffer profiles, MTU consistency, priority mapping, congestion control, FEC compatibility, NIC and driver support, failure recovery, monitoring, and telemetry.

A fabric can have enough raw bandwidth and still perform poorly if congestion control and queue behavior are not aligned across servers, switches, and optics. Use the AI Cluster Lossless Networking with RoCEv2 guide for the detailed PFC, ECN, buffer, and GPUDirect planning considerations.

A sizing or quotation request should include:

  • Total GPU count
  • GPU model and GPUs per server
  • Number of GPU servers
  • NIC model and port count
  • 400G or 800G target speed
  • Storage system and interface speed
  • Management node count
  • Target Leaf-Spine topology
  • Required redundancy
  • Fiber distance
  • DAC/AOC length
  • Breakout requirements
  • Planned expansion size
  • Optics and cable spares
  • Power and airflow restrictions

Frequently Asked Questions

Q1 How many switches are needed for a 1,000-GPU cluster?
There is no universal answer. Cisco’s reference architecture shows a 1,024-GPU, 128-node three-tier compute design with 16 leaf switches, 16 spine switches, and 8 Super-Spine switches. A different server NIC layout, port density, or topology can produce a different result.
Q2 Is one switch required for every GPU server?
No. Multiple GPU servers connect to each leaf switch according to the server port count, link speed, rail design, and available downlink ports.
Q3 Does an AI cluster need a separate storage network?
Not always, but storage should be planned separately even when it is physically converged. Dataset access, checkpoint traffic, and parallel storage may require different bandwidth and congestion policies from the GPU compute fabric.
Q4 Is 800G always better than 400G?
No. 800G provides higher bandwidth per physical port, but it may increase optic, cabling, compatibility, and deployment complexity. The correct choice depends on server interfaces, switch radix, target scale, and growth requirements.
Q5 How many optical transceivers should be included?
Count every physical link endpoint. A redundant Leaf-Spine design requires more modules than a single-link design, and breakout configurations require a separate count for each resulting lane or endpoint.
Q6 Should management switches be included in the GPU cluster BOM?
Yes. BMC, switch management, provisioning, monitoring, console, and PDU connectivity are operationally necessary even though they do not carry the main GPU traffic.
Q7 What is the first thing to confirm before choosing a switch?
Confirm the GPU server configuration and its network interface layout. The number of GPUs alone is not enough to determine leaf port demand or switch count.

Final Takeaway

AI GPU cluster switch sizing begins with the server and network topology, not with a simple GPU-to-switch ratio.

For a reliable 400G or 800G BOM, calculate GPU servers, GPU-facing links, Compute Leaf and Spine layers, Super-Spine requirements, storage and management networks, optical modules, breakout assemblies, DAC/AOC and fiber cables, redundant links, and spares.

Cisco’s 8000 Cloud Reference Architecture provides useful examples from 1K to 32K GPUs, but those figures apply to a defined Cisco and NVIDIA design. For a different GPU server, switch platform, NOS, or topology, the BOM must be recalculated.

If you are preparing a cluster purchase, send the GPU count, server configuration, target link speed, storage requirements, and expansion plan for a switch, optics, and cabling review.