Networks for neocloud and GPU cloud operators

IP Infusion delivers the networks a GPU cloud runs on: OcNOS-DC on qualified switches for the lossless GPU fabric, the storage network, the tenant front-end and out-of-band management, under one support contract.

Book an architecture review Get the OcNOS-DC datasheet
64 × 800GTomahawk 5 switches for the GPU fabric
16,384GPUs in the largest reference design
600+operators run OcNOS

One GPU server, four networks

A GPU cloud connects every server to four networks: the GPU back-end fabric, storage, the tenant front-end and management. OcNOS-DC runs all four, so one team can operate one software family across the cloud.

Pick a network to see what it carries

A GPU server connects to four networks: the GPU back-end fabric, storage, the tenant front-end and out-of-band management Spine Spine Rail leaf Rail leaf GPU back-end fabric Storage leaf Storage Tenants Front-end leaf Tenant front-end Management Out-of-band, 1G GPU server GPU NICs Storage NIC Front-end NIC BMC

GPU back-end. GPU collectives run on their own routed fabric. PFC and ECN over Layer 3 keep RoCEv2 traffic lossless, and dynamic load balancing moves flowlets to less-loaded uplinks. Tomahawk 5 switches carry 64 × 800G.

Illustrative.

Training and inference ask different things of the network

Neoclouds serve both. Training fills a few large clusters with steady, synchronized traffic. Inference traffic spans many sites and arrives in bursts.

Training vs inference, across the network metrics that shape the fabric.
What the network sees Trainingthe AI factory Inferencethe real-time service
FootprintA few large GPU clustersMany regional and edge sites
TrafficBulk, synchronized GPU-to-GPU collectivesSmall, latency-critical requests
Load profileSustained, near full GPU utilization for long runsBursty and elastic, spikes in seconds
The network has to beLossless under sustained bandwidthLow-latency and redundant at every layer

What OcNOS-DC runs in a GPU cloud

The features each network relies on, from the OcNOS Feature Matrix.

CapabilityWhat it doesWhere
PFC and ECN over Layer 3Lossless RoCEv2 across a routed fabric, with ETS between traffic classesAll OcNOS-DC
Dynamic load balancingMoves flowlets to less-loaded uplinksSelected switches
PFC deadlock detection and recoveryDetects a stalled priority queue and recovers itSelected switches
EVPN-VXLANLayer 2 EVPN over VXLAN, with a VRF and VNIs per tenantTrident 3 and 4
Tenant limitsMAC limits per tenant and route-target filteringSelected switches
PFC and ECN over VXLANLossless storage traffic on an EVPN-VXLAN fabricSelected switches
MLAG and EVPN multihomingMLAG in active/active mode; EVPN in single-active or port-active modeSelected switches
Coherent optics in the switch400G and 800G ZR and OpenZR+ links between sitesSelected switches
Management VRF and ZTPOut-of-band management and zero touch provisioningAll OcNOS-DC

Support varies by switch model and license. Check the OcNOS Feature Matrix.

Switches by network

Each network uses a switch suited to its traffic, all running OcNOS-DC.

Edgecore AIS800-64D switch

GPU fabric spine and leaf

Edgecore AIS800-64D

Tomahawk 5 · 51.2 Tbps

64 × 800G

Edgecore AS9736-64D switch

400G GPU rail leaf

Edgecore AS9736-64D

Tomahawk 4 · 25.6 Tbps

64 × 400G

UfiSpace S9300-32D switch

Front-end and storage leaf

UfiSpace S9300-32D

Trident 4 · 12.8 Tbps

32 × 400G

UfiSpace S6301-56ST switch

Out-of-band management

UfiSpace S6301-56ST

Trident 3 · 120 Gbps

48 × 1G, 8 × 10G

OcNOS on open switches carries the AI fabric at 40 to 60% lower hardware cost than proprietary platforms at comparable port speeds and radix. See how the lossless AI fabric works. See the full list of qualified switches on the OcNOS Hardware Compatibility List.

Operators running OcNOS in the data center

IzarLink

An EVPN-VXLAN core that aggregates 25G access into a redundant 100G core and interconnects its data centers and peering locations.

Read the story →

Nextlink Internet

OcNOS-DC for core data center operations and OcNOS-SP for middle-mile transport, on Edgecore hardware across 11 states.

Read the story →

Scott Data

OcNOS runs on UfiSpace, Edgecore and Celestica hardware across a Tier III certified data center campus with colocation services.

Read the story →

“IP Infusion OcNOS provided the technical depth and improved network economics that enabled us to timely upgrade our data center core with future-proof scale, security, and resilience.”

Alexis Guirlé, CEO of IzarLink

Questions about neocloud networking.

What network do neocloud and GPU cloud providers need for AI?
Four networks: a routed, lossless back-end fabric for GPU collectives, a storage network, a front-end network where customers reach their GPUs, and out-of-band management. OcNOS-DC runs all four on qualified switches.
Can one platform support both AI training and inference?
Yes. The same OcNOS-DC family runs the lossless GPU fabric that training needs and the front-end and storage networks that inference uses, so one team runs one software stack across the cloud. Switches are sized per role.
How do GPU clouds keep customers isolated?
On the front-end network, EVPN-VXLAN gives each tenant its own VRF and VNIs, with MAC limits per tenant and route-target filtering. These features run on Trident 3 and Trident 4 switches; support varies by switch model, so check the OcNOS Feature Matrix.
How does the fabric keep inference available when a link or switch fails?
BFD detects link failures, and BGP ECMP moves traffic to the remaining paths. MLAG connects servers to a pair of switches in active/active mode. On EVPN-VXLAN front-end leaves, EVPN multihoming in single-active or port-active mode connects servers to two leaves.
How do I evaluate network vendors for AI training and inference?
Check that each switch is qualified for the software you plan to run, for each model and ASIC stepping. Confirm that the features you need are listed for that switch, and find out who supports the switch and the software. IP Infusion publishes an OcNOS Feature Matrix for each platform and supports switch and software under one contract.
How does a neocloud connect its sites?
Supported OcNOS-DC switches support coherent optics in their ports: 400G ZR and OpenZR+ on the AS9736-64D, AS9726-32DB and S9300-32D; 800G ZR and OpenZR+ on the AIS800-64D and S9321-64E. For longer links, see IP over DWDM on OcNOS-SP routers.

Datasheet and solution briefs.

The OcNOS-DC datasheet and two briefs to share with your team. Quick form, and the PDF downloads immediately.

Solution Brief

OcNOS 800G Lossless AI Fabric

Lossless RoCEv2 fabric on Tomahawk 5 spines: SKU tiers, qualified platforms, and deployment architecture.

Get the brief
Solution Brief

EVPN-VXLAN Data Center Fabric

Leaf-spine data center fabric on EVPN-VXLAN: a VRF and VNIs per tenant, EVPN multihoming, BGP unnumbered and route-target filtering.

Get the brief
Datasheet

OcNOS-DC Datasheet

Full OcNOS-DC specification: the EVPN-VXLAN and Ethernet for AI feature set, software SKUs, supported hardware platforms, and the solution ordering guide.

Get the datasheet

Plan your GPU cloud network.

Tell us your GPU scale, tenants and sites. An IP Infusion engineer will map each network to qualified switches.

Book an architecture review Open the AI fabric design suite