800GbE NIC Deployment Calculator for AI Servers

800G endpoint and port planning

800GbE NIC Deployment Calculator for AI Servers

An 800GbE deployment can run out of switch ports, PCIe slots or optics long before it runs out of theoretical bandwidth. This calculator converts server count and NIC density into physical adapters, network ports, aggregate terabits and a first-pass switch count. It works for NVIDIA ConnectX-9, Broadcom Thor Ultra or another 800G-class NIC because the arithmetic is kept hardware-neutral.

Interactive calculator

800GbE NIC Deployment Calculator

Enter your own topology and bandwidth assumptions. These planning calculators estimate raw links, capacity and ratios; they do not certify fabric goodput, rail mapping, cable reach, firmware interoperability, electrical design, cooling or OEM compatibility.

Live Amazon supporting hardware

Current supporting components and Price Options

Compare current Amazon listings relevant to this guide. Product availability and prices can change.

Loading current Amazon listings...

Quick answer

What should be counted in an 800GbE deployment?

Count NICs, physical network ports, host PCIe capacity, switch-facing links, optics or DAC/AOC assemblies, redundant capacity and the uplink budget. A server with two 800G NICs contributes 1.6 Tb/s of raw endpoint bandwidth and two physical 800G links unless the selected adapter or breakout changes that mapping.

Physical compatibility is the gating step after arithmetic

The calculator can tell you that a design needs hundreds of adapters and links, but it cannot prove that the target server has enough PCIe Gen6 lanes, airflow, power or qualified cable reach. Validate the exact NIC SKU, host slot arrangement and switch port mode before converting counts into purchase orders.

800GbE deployment checklist

High-speed NIC projects fail when line-rate arithmetic is done without host and physical-layer checks.

Physical items to verify in an 800GbE NIC deployment.
ItemWhat to countTypical constraintVerification
NIC cardsNodes × adapters per nodePCIe slots and airflowServer slot map
Network portsNICs × ports per NICSwitch radixAdapter OPN / port mode
Host bandwidthConfigured NIC line ratePCIe generation and lanesServer BIOS / CPU topology
Optics / cablesOne path per physical linkReach and connectorQualification matrix
Switch portsEndpoint links plus uplinksBreakout and oversubscriptionSwitch config
SparesAdapters, ports and mediaFailure and growth policyOperations standard

Before you use the result for procurement

Draw the physical topology

Map every server-facing port, leaf uplink, spine link and network plane. Aggregate bandwidth alone can hide impossible port or lane assumptions.

Verify exact endpoints

Confirm NIC form factor, PCIe generation, host lane budget, port speed, connector and firmware support on the exact server platform.

Qualify optics and cables

Match OSFP/QSFP form factor, lane rate, breakout, reach, fiber type and both endpoint qualification lists. Do not treat equal headline speed as automatic compatibility.

Test failure and congestion behavior

Validate oversubscription, ECMP or multiplane path behavior, switch failure domains and recovery under the traffic patterns the AI workload will actually generate.

Begin with adapter count, not aggregate terabits

Multiply server nodes by NICs per node to get the physical card count. This sounds obvious, but rack-scale systems can expose network bandwidth through integrated or specialized configurations that do not map one-to-one to an ordinary add-in card. A hardware-neutral deployment should state the actual adapter model and physical quantity.

Once card count is known, procurement can check slot inventory, bracket type, power and airflow. Those constraints are more concrete than saying the cluster needs “200 Tb/s of networking.”

Port count can differ from NIC count

Some adapters have one high-speed cage while others provide multiple lower-rate ports. ConnectX-9, for example, has different OSFP and QSFP112 models. The calculator therefore asks for physical ports per NIC rather than assuming one.

Port count drives switch attachments and media. A two-port adapter running two 400G links is not the same cable plan as a single 800G OSFP adapter even if the raw aggregate rate is similar.

Host PCIe bandwidth must be validated

Both ConnectX-9 and Thor Ultra use PCIe Gen6 x16-class host interfaces for their 800G designs. Older slots may physically accept a card while limiting available host bandwidth or failing vendor qualification.

Inspect CPU lane topology and BIOS settings, not just the slot label. In multi-socket systems, placing the NIC behind the wrong CPU or switch can add NUMA penalties that show up as lower GPU-to-network throughput.

800G creates a serious airflow requirement

NVIDIA explicitly states that ConnectX-9 is designed for data-center servers with proper power and airflow and is not intended for ordinary desktop or workstation installation. Similar thermal caution applies to any dense 800G adapter.

A high-speed NIC can be electrically correct and still throttle or fail if the chassis cannot remove heat from the adapter and optic. Confirm supported airflow direction, heatsink type and slot spacing with the server vendor.

Switch radix converts endpoint ports into chassis count

Divide endpoint-facing physical links by the number of switch ports that are actually available for those links. The “usable” switch port input should exclude ports reserved for uplinks, spares or another rail.

If the topology uses dedicated planes or rails, calculate each domain separately. Pooling all ports into one theoretical switch can hide the physical isolation rules required by the reference architecture.

Spare capacity belongs in the initial purchase model

A fabric with every port occupied has little room for replacement, incremental racks or topology changes. The calculator reduces usable capacity by the entered spare percentage before estimating switch count.

Operational spare policy should also cover NIC cards and optics. High-speed components may have longer replacement lead times than common 10G or 25G equipment, making on-site spares economically sensible.

Cable type is dictated by distance and qualification

Short 800G links may use direct-attach copper or active cables when supported, while longer distances require optics and fiber. The price, power and service model can differ sharply across those choices.

Measure rack and row distances before ordering media. A cable selected from connector names alone can fail because the lane mapping, reach or endpoint firmware does not support that exact assembly.

Breakouts alter both speed and port math

An 800G physical port can sometimes be broken into several lower-speed lanes, but the supported breakout modes depend on the switch, NIC, optic and software. Breakout can improve radix for multiplane designs or mixed-speed clusters.

Do not count theoretical electrical lanes as usable ports until the exact hardware manual confirms the mode. A spreadsheet that assumes every 800G cage can become four independent 200G links may materially understate switch requirements.

Aggregate bandwidth should be expressed in both directions carefully

Network vendors quote line rate using standard conventions, while application planners may speak about send plus receive or full-duplex traffic differently. The calculator reports raw configured endpoint line rate without inventing a doubled bidirectional number.

When comparing to PCIe or memory bandwidth, make sure both sides use the same direction convention and units. Many apparent bottlenecks are unit or accounting mismatches rather than real hardware limits.

Redundancy changes normal utilization

If a network must survive a NIC or switch failure without throttling workloads, normal operation should leave enough capacity for failover traffic. A design that runs at 95% line rate before a fault has very little recovery room.

Use the spare percentage and topology-level plane redundancy together. Spares handle physical capacity and maintenance; plane or rail architecture determines whether alternate paths actually exist.

Retail listings are best for lab and reference pricing

Cloudzat can surface available NVIDIA, Broadcom, optic and cable listings from Amazon, but the latest 800G enterprise parts may be absent, used, OEM-only or sold through specialized channels. Product title evidence is not a substitute for vendor qualification.

Use the table to understand component categories and market price ranges, then obtain enterprise quotes for exact OPNs, support coverage and volume purchases.

Acceptance testing should saturate the real data path

A NIC can pass link tests while GPU-to-GPU throughput remains constrained by PCIe placement, peer-memory configuration, IOMMU settings, NUMA topology or switch congestion. Run traffic that starts and ends where the production AI workload will.

Measure throughput, tail latency, CPU overhead, retransmissions and thermal behavior over sustained runs. The goal is to prove the server, NIC, cable and switch as one path, not merely to confirm an 800G link state.

Methodology and sources

The calculator is intentionally vendor-neutral and counts physical adapters, ports and links from the user’s inputs. Public ConnectX-9 and Thor Ultra documentation is used only to explain why PCIe Gen6, airflow and qualification matter; the arithmetic does not assume a specific NIC.

As an Amazon Associate, Cloudzat may earn from qualifying purchases. Marketplace listings on these pages are supporting networking hardware such as NICs, switches, optics and high-speed cables. A marketplace row is not represented as a Spectrum-6 switch, ConnectX-9 SuperNIC, Thor Ultra NIC or qualified NVIDIA fabric unless the exact listing evidence supports that identity. Verify model, speed, connector, firmware, warranty and OEM qualification before purchase.

Frequently asked questions

How many 800G NICs do I need?

Multiply the number of server nodes by the intended NICs per node, then add operational spares if required.

How many switch ports does one 800G NIC use?

Usually one physical link per active network port, but multi-port adapters and breakout modes can change the mapping. Verify the exact NIC SKU.

Does an 800G NIC need PCIe Gen6?

Current leading 800G AI NICs such as ConnectX-9 and Broadcom Thor Ultra use PCIe Gen6 x16. The host must be validated for the exact adapter.

Can an 800G port break out to 4 x 200G?

Some switch and optic combinations support breakouts, but it is not universal. Check the exact switch, NIC and cable documentation.

Should I use DAC or optics?

Use the shortest qualified medium that meets reach, power and service requirements. Passive DAC can be economical at short distances; optics are needed for longer runs.

How much bandwidth is two 800G NICs?

They provide 1.6 Tb/s of raw configured line rate in the same direction convention, before protocol and application overhead.

Why reserve spare switch ports?

Spare ports make failures, growth and cable moves easier and prevent a small topology change from forcing another switch immediately.

Can I buy ConnectX-9 or Thor Ultra on Amazon?

Availability varies. Verify exact model, condition and seller; enterprise channels may be more appropriate for current 800G parts.

What server checks are required?

Verify PCIe generation and lanes, slot location, power, airflow, bracket/form factor, BIOS and vendor support.

What performance test should I run?

Use GPU/XPU-aware network benchmarks and the target collective framework, not only host TCP throughput, then test sustained load and failure behavior.

Scroll to Top