Switches for AI fabrics – 400G, 800G and InfiniBand
We design and deliver the network for GPU clusters: Dell PowerSwitch and NVIDIA switches, Enterprise SONiC, RoCEv2 and InfiniBand – from the spine/leaf topology to commissioning.
-
400G or 800G Ethernet – PowerSwitch Z and SN series
-
InfiniBand Quantum-X800 – up to 115.2 Tbps
-
Enterprise SONiC – an open network OS
Who it is for
For companies building or expanding AI infrastructure
An AI cluster has a different network profile than an office: traffic is concentrated between GPU nodes and storage, in large blocks, and sensitive to drops.
GPU clusters and private AI
A network carrying traffic between GPU nodes and storage – without blocking and without losses on RDMA traffic.
Data centers and server rooms
Spine/leaf topologies with 400G uplinks and 800G headroom, plus a management network outside the fabric.
Upgrading an existing network
Moving from 1G/10G to 400G in stages, following an agreed migration plan and switching windows.
HPC and research teams
InfiniBand or Ethernet depending on the workload profile – we help choose, design and commission it.
Benefits
Why AI traffic needs different switches
Training and inference carry parallelized traffic between hundreds of GPUs – every delay or drop costs the whole cluster time.
A fabric without blocking
Capacities of 25.6 to 51.2 Tbps per switch and line-rate forwarding – traffic between GPU nodes never waits on port oversubscription.
400G and 800G today
OSFP and QSFP-DD ports are multi-rate: 100G, 200G, 400G and 800G in the same fabric – the network grows with the GPU nodes, without a platform swap.
RoCEv2 and lossless transport
Lossless Ethernet for RDMA: PFC and ECN keep GPU ↔ storage traffic loss-free – training runs fast and with stable steps.
Enterprise SONiC
The open network OS (SONiC) in Dell’s Enterprise distribution – telemetry, controlled updates and support, without lock-in to closed software.
NVIDIA Spectrum-X
The SN5600D carries Spectrum-4 silicon: adaptive routing and congestion control developed precisely for AI traffic.
Our part of the job
Topology design, optics sizing, mounting, configuration and commissioning – with documentation for your IT team.
Dell AI switch catalog
Models on offer – Dell PowerSwitch and NVIDIA
400G and 800G Ethernet fabrics on Enterprise SONiC and InfiniBand for the most demanding clusters – every card goes straight to inquiry, sized by the number of GPU nodes.
Dell Pro Precision
Dell PowerSwitch Z9864F-ON
800G Ethernet mreža za najveće GPU klastere
- 64× 800 GbE OSFP112 + 2× SFP+ porta
- Broadcom Tomahawk 5 – 51,2 Tbps, 20,3 Bpps
- Enterprise SONiC: RoCEv2, ECN i dinamičko balansiranje
Spine ili leaf sloj 800G mreže – rezerva za sledeću generaciju GPU čvorova.
Dell Pro Precision
Dell PowerSwitch SN5600D
NVIDIA Spectrum-X leaf switch za GPU klastere
- 64× 800 GbE OSFP + 1× 25 GbE SFP28
- NVIDIA Spectrum-4 – 51,2 Tbps, 33,3 Bpps
- Adaptivni routing i congestion control za AI saobraćaj
Deo NVIDIA Spectrum-X ekosistema – prirodan par uz GPU sisteme iz naše AI ponude.
Dell Pro Precision
Dell PowerSwitch Z9664F-ON
400G Ethernet mreža za spine i leaf slojeve
- 64× 400 GbE QSFP56-DD + 2× 10 GbE SFP+
- Broadcom Tomahawk 4 – 25,6 Tbps, 10,2 Bpps
- Multi-rate portovi: 100G/200G/400G u istoj mreži
Najčešći izbor za prvu 400G mrežu – ulazak u AI klasu bez 800G budžeta.
Dell Pro Precision
Dell PowerSwitch SN5400
400G leaf switch za GPU i skladišne slojeve
- 64× 400 GbE QSFP-DD + 2× 25 GbE SFP28
- NVIDIA Spectrum-4 – 25,6 Tbps, latencija 620 ns
- Enterprise SONiC – RoCEv2 za RDMA saobraćaj
Niska latencija (620 ns) – za saobraćaj GPU ↔ skladište.
NVIDIA
NVIDIA Quantum-X800 Q3400-RA
InfiniBand mreža za najveće HPC i AI klastere
- 144× 800 Gb/s kroz 72 OSFP kaveza – 115,2 Tbps
- NVIDIA Quantum-3 silicijum, SHARP in-network computing
- Radi uz NVIDIA GPU sisteme iz naše AI ponude
Za klastere gde je InfiniBand standard – ista optika i rack logika.
Not sure which network your cluster needs?
Tell us how many GPU nodes the network carries and which applications run on it – we will propose a topology, models and optics without overdoing it.
What is useful to include
- The number of GPU nodes and the NIC per node
- Ethernet or InfiniBand – what you already run
- The deadline for phase one and the expansion plan
No models match this combination. Deselect filters or contact us.
For the out-of-band management network we propose the Dell PowerSwitch SN2201 and S3248T-ON from the same catalog. Optics and DAC/AOC cables come with the project – sized by the rack layout and link lengths.
FAQ
Frequently asked questions
Solutions
Related solutions
Solutions that most often go together with this project.
Enterprise Networks & Wi-Fi
Network design and build: switches and Wi-Fi 6/7 access points, installation and commissioning. Reliable internet in the office, warehouse and production.
Servers & storage
Rack servers, all-flash and hybrid storage from Dell, HPE, Lenovo and Supermicro: design, rack, installation, commissioning and virtualization by agreement.
Network security
Firewall, UTM and VPN protection for companies: control of traffic, access and remote work. Appliance, licence and commissioning.
Request an AI network specification
Tell us the number of GPU nodes, the applications and the deadline – we will come back with a topology, models and optics proposal.
Planning a GPU cluster or a network upgrade?
The network is easiest to design before the GPU nodes reach the rack. Invite us into the early stage – with the optics, timelines and documentation.
