Broadcom Tomahawk 6: 1.6T Ethernet for AI fabrics.

Tomahawk 6 is Broadcom's StrataXGS Ethernet switch chip for AI and data center fabrics.

102.4 Tbpsswitching capacity on one chip, per Broadcom
64 x 1.6TbEmaximum port configuration
512 x 200GbEmaximum radix configuration, per Broadcom

OcNOS on Tomahawk 6

Coming soon. OcNOS-DC support for Tomahawk 6 is planned for the upcoming OcNOS 7.2 release.

Tomahawk 4 and Tomahawk 5 AI fabrics run on OcNOS-DC in production today.

Planning a 1.6T deployment? Contact us about platform availability and migration planning.

Compare with Tomahawk 5 What OcNOS adds
Against Tomahawk 5

What changes from Tomahawk 5?

At 200GbE, Tomahawk 5 offers 256 ports and Tomahawk 6 offers 512, so a Tomahawk 6 switch can connect twice as many 200G endpoints. That can reduce the number of switches a given fabric design needs. Actual switch count, cabling, power and cost depend on the topology, oversubscription, port mix and cluster architecture.

Tomahawk 5 against Tomahawk 6. Chip figures are Broadcom's.
Chip level Tomahawk 5 Tomahawk 6
Switching capacity 51.2 Tbps 102.4 Tbps
Maximum port speed 64 x 800GbE 64 x 1.6TbE
Radix at 200GbE 256 ports 512 ports
Process 5 nm 3 nm
Topologies Clos and non-Clos, including torus, Dragonfly, Dragonfly+ and Megafly Arbitrary topologies, including scale-up, Clos, rail-optimized, rail-only and torus
OcNOS-DC Available today Planned, OcNOS 7.2

Technical silicon detail: devices, SerDes, congestion handling and Ultra Ethernet.

In the fabric

What 1.6T means for an AI cluster.

If your endpoints are 200G

Higher radix can raise the number of endpoints attached to each switch. How to size a switch by radix is on the AI networking switches page.

If your endpoints are 400G or 800G

You do not have to replace every endpoint to move to a 1.6T-capable switch. The speed a switch port can run is separate from the speed each endpoint uses.

If you are designing a new large cluster

Higher bandwidth and radix can change how many leaf and spine switches, optics, cables and ports the design needs. Links that move to 1.6T need optics and cabling rated for it; Broadcom cites the 200G SerDes for the longest reach over passive copper.

If you run an OcNOS fabric today

The software you run today, and what is planned for Tomahawk 6, are under What OcNOS adds.

Send us your cluster size and endpoint speed for a fabric design review.

Technical detail

Silicon, congestion handling and Ultra Ethernet.

The Tomahawk 6 family includes two conventional switch devices and a co-packaged optics variant.

Tomahawk 6 devices. Every figure is Broadcom's.
Device Family name SerDes Main distinction Status
BCM78910 Tomahawk 6 128 Peregrine cores, 1024 lanes of 100G PAM4 Twice the lane count of the BCM78914, at half the lane rate Broadcom-stated: production volume for the Tomahawk 6 family switch series, announced March 12, 2026
BCM78914 Tomahawk 6 64 Condor cores, 512 lanes of 200G PAM4 Broadcom cites the 200G SerDes for the longest reach over passive copper interconnect Broadcom-stated: production volume for the Tomahawk 6 family switch series, announced March 12, 2026
BCM78919 Tomahawk 6-Davisson 64 Condor 3 nm cores, 200G Co-packaged optics with 16 x 6.4 Tbps optical engines and a 267 MB unified packet buffer Broadcom-stated: sampling; the product page reads limited release

Cognitive Routing 2.0

Broadcom-stated: Cognitive Routing 2.0 on Tomahawk 6 combines advanced telemetry, dynamic congestion control, rapid failure detection and packet trimming, with hardware-based link failover in the silicon. How the lossless path works is on the RoCEv2 page.

Ultra Ethernet

Broadcom-stated: Tomahawk 6 is Ultra Ethernet Consortium compliant. IP Infusion is a member of the Ultra Ethernet Consortium. What the transport changes is on the Ultra Ethernet page.

OcNOS roadmap

Planned: packet trimming, fabric-wide global load balancing, latency-based ECN, link-level retry and credit-based flow scheduling, set out on the AI fabric architecture page.

Scale-out and scale-up

Where Tomahawk 6 sits in an AI data center.

Tomahawk 6 can sit in both domains. The boundary between the two is on the GPU fabric page.

Broadcom Tomahawk 6 in two roles in one AI cluster. Scale-out, between racks: two Tomahawk 6 spines over four Tomahawk 6 leaves, each leaf one color-coded rail, with every rack sending one NIC to each rail; labeled OcNOS-DC planned, RoCEv2, eBGP and 1.6T. Scale-up, inside the opened rack: one Tomahawk 6 switch connecting eight XPUs directly over Broadcom Scale-Up Ethernet.
Scale-out between racks, scale-up inside the rack. The planned OcNOS-DC support is for the scale-out fabric.
Silicon and software

What OcNOS adds on top of the switch silicon.

OcNOS runs on top of the switch silicon. It runs AI fabrics in production today on Tomahawk 5 at 800G and Tomahawk 4 at 400G, on platforms listed in the hardware compatibility list.

OcNOS software

OcNOS-DC, the data center edition

Forwarding
eBGP unnumbered underlay with multipath load sharing
Congestion control
PFC and ECN for lossless RoCEv2
Load balancing
Dynamic load balancing across fabric links
Telemetry
gNMI streaming telemetry with OpenConfig models
Failure recovery
BFD, and PFC deadlock detection and recovery
Automation
Zero-touch provisioning, NETCONF with YANG, Ansible, and one CLI on every HCL-listed data center platform

Broadcom Tomahawk 6 silicon

StrataXGS Ethernet switch chip

Capacity and ports
102.4 Tbps; 64 x 1.6TbE or 512 x 200GbE
In the silicon
Cognitive Routing 2.0 and Ultra Ethernet Consortium compliance, per Broadcom

Planned for Tomahawk 6 in OcNOS 7.2

  • One OcNOS image that will span Tomahawk 4, Tomahawk 5 and Tomahawk 6
  • One CLI and one automation toolchain
  • One routed underlay design that will span 400G, 800G and 1.6T

Supported Tomahawk 6 platforms will be announced with the release. Feature detail by release is in the OcNOS feature matrix; contact our sales team for Tomahawk 6 plans.

Why IP Infusion

Built by many. Delivered by one.

Broadcom supplies the switching silicon and hardware partners build the platforms. IP Infusion delivers the complete system: OcNOS software on open hardware it qualifies per model, optics with proven interoperability, and support and RMA from one vendor.

Named at the Tomahawk 6 launch

Broadcom named IP Infusion in its Tomahawk 6 launch announcement of June 2025. IP Infusion maintains expertise with the Broadcom StrataXGS product line.

One portfolio covers the data center and the service provider network

OcNOS-DC runs the AI back end, storage, front end and out-of-band management networks in the data center, and OcNOS-SP runs the service provider transport beside them. One vendor and one support relationship cover both.

More than one place to buy the hardware

HCL-listed Tomahawk 5 platforms come from two hardware makers, so an 800G fabric can be built on a dual-source bill of materials.

Tomahawk 6 questions

What is Broadcom Tomahawk 6?
Tomahawk 6 is Broadcom StrataXGS switch silicon for data center and AI Ethernet fabrics and, according to Broadcom, the first single chip to carry 102.4 Tbps. Broadcom builds it as two parts, both on a 3 nm process: the BCM78910 with 128 Peregrine 106.25G PAM4 SerDes cores, which is 1024 lanes of 100G, and the BCM78914 with 64 Condor 212.5G PAM4 SerDes cores, which is 512 lanes of 200G. Broadcom lists configurations of up to 64 x 1.6TbE, or 512 ports of 200GbE, on a single chip.
Can I run OcNOS on Tomahawk 6?
Coming soon. OcNOS-DC support for Tomahawk 6 is planned for the upcoming OcNOS 7.2 release. Supported Tomahawk 6 platforms will be announced with that release; contact the IP Infusion sales team for details. Available today: OcNOS-DC runs AI fabrics in production on HCL-listed Tomahawk 5 hardware at 800G and Tomahawk 4 hardware at 400G, with the same industry-standard CLI and automation it will carry to Tomahawk 6. The hardware compatibility list is the record of every platform it runs on.
How does Tomahawk 6 compare with Tomahawk 5?
Broadcom describes Tomahawk 6 as doubling the throughput of Tomahawk 5. Its published figures give 102.4 Tbps against 51.2 Tbps, 3 nm against 5 nm, 512 ports of 200GbE against 256, and a top port rate of 1.6TbE against 800GbE at the same 64 ports. At the same endpoint speed, higher switching capacity can reduce the number of switches a fabric design needs. For the 800G part itself, see Tomahawk 5; for the whole portfolio, see the Broadcom silicon hub.
What changes when an AI fabric moves from 800G to 1.6T?
Capacity: at the same endpoint speed, higher switching capacity can reduce the number of switches and tiers a fabric design needs; Broadcom publishes 100,000 or more XPUs in a two-tier scale-out network at 200 Gbps per link, and LightCounting, quoted in Broadcom's March 2026 announcement, puts a 128K-XPU network at two switch tiers. Topologies: Broadcom adds scale-up, rail-optimized and rail-only designs to the Tomahawk 6 list. Congestion: Broadcom lists rapid failure detection and packet trimming under Cognitive Routing 2.0, and hardware-based link failover. Transport: Broadcom lists Ultra Ethernet Consortium compliance as a product feature. How these pieces fit together is on the what is an AI fabric page.
What is the difference between scale-out and scale-up on Tomahawk 6?
Scale-out is the Ethernet fabric between racks, the Clos that carries collectives from leaf to spine and back. Scale-up is the domain inside one rack, where accelerators reach each other directly. Broadcom states that Tomahawk 6 is optimized for both, publishes a scale-up cluster size of 512 XPUs, and has released the Scale-Up Ethernet framework as an open specification for the accelerator side of that link. On Tomahawk 6, OcNOS-DC is planned for the scale-out fabric. The AI fabric architecture page shows where each one sits.
Is Tomahawk 6 Ultra Ethernet compliant?
Yes, according to Broadcom, which lists Ultra Ethernet compliance as a feature on the BCM78910 series product page and in the Tomahawk 6 launch release. IP Infusion is a member of the Ultra Ethernet Consortium. What the transport itself changes is on the Ultra Ethernet page.
What can I buy from IP Infusion for an AI fabric today?
OcNOS-DC on HCL-listed Tomahawk 4 hardware at 400G and HCL-listed Tomahawk 5 hardware at 800G, delivered as a complete system: the software, the open hardware, the optics interoperability and one support contract. The mechanisms an AI back end needs are in that software now, including lossless RoCEv2 with priority flow control and ECN, dynamic load balancing, PFC deadlock detection and recovery, and gNMI streaming telemetry. The AI fabric solution page covers the whole offer.
Do 800G and 400G endpoints still attach to a 1.6T switch?
Yes. Broadcom lists Tomahawk 6 configurations using 400GbE and 800GbE as well as 1.6TbE, so a fabric can keep existing 400G and 800G connectivity rather than moving every endpoint to 1.6T at once. The speed a switch port can run is separate from the speed each endpoint uses. Broadcom cites the 200G SerDes for the longest reach over passive copper interconnect.

Planning a 1.6T fabric?

Send us the size of the cluster and the endpoint speeds, and we will review the design and the platform options with you.

Contact us How to buy OcNOS feature matrix
Datasheet and Solution Briefs

AI fabric downloads

The product datasheet plus short, technical downloads that go further than this page.

Reference Design

AI Fabric Reference Design

A worked 1,024-GPU rail-optimized fabric on OcNOS-DC and Tomahawk 5: 32 leaves, 16 spines, 1:1 non-blocking, and where a super-spine tier begins.

Get the design
Solution Brief

OcNOS 800G Lossless AI Fabric

Non-blocking RoCEv2 fabric on Broadcom Tomahawk 4/5 spines: SKU tiers, validated platforms, and deployment architecture.

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

Broadcom, StrataXGS, StrataDNX, Tomahawk, Jericho, Qumran and the BCM part numbers referenced on this page are trademarks of Broadcom Inc. and its affiliates. Ultra Ethernet and Ultra Ethernet Consortium are marks of the Ultra Ethernet Consortium. All are used here for identification only. IP Infusion is an independent company and is not affiliated with, endorsed by or sponsored by these companies. Chip figures on this page are those published by Broadcom, attributed where they appear and subject to change.