Scale out

Many rows become one AI cluster

Automotive Ethernet Solutions

A scale-out network wires thousands of racks — row upon row across a floor — into a single AI cluster, tied together by hundreds of thousands of optical links. Marvell builds the interconnect silicon and switching at the heart of that fabric, matched to every link within it.

To build a modern AI cluster, row upon row of racks have to be wired together into one fabric — thousands of them, under one roof, moving data east to west between accelerators at the heart of every training run and every large inference job. That fabric is scale out. It sits between scale up, which tightly couples accelerators into a single compute domain, and scale across, which connects separate data centers. Scale out is the tier that turns a room full of racks into one machine.

AI data center Scale out connectivity

 

Scale out is the connectivity that links racks across many rows into a single AI cluster within a data center. It’s the east-west fabric that lets thousands of accelerators work as one.

Scale-out networking demands more than one kind of connection


A scale-out network, or fabric, ties thousands of racks — row upon row, under one roof — into a single AI cluster. Across that fabric, hundreds of thousands of links do the work, but they’re not carbon copies of one another. A short hop between adjacent racks and a reach across a data hall place entirely different demands on the connection. Each link balances five variables against one another:

  • Bandwidth: how much data per second the link must move
  • Latency: how much delay the workload can absorb
  • Power: watts per bit, increasingly the constraint that caps how far a buildout can expand
  • Cost: capex per link, multiplied by hundreds of thousands
  • Distance: from a couple of meters between racks to 500m+ across a floor

No single interconnect type optimizes all five. Reach for maximum distance and you spend power and cost; strip out power and cost and you give up reach. A fabric that scales efficiently is the one that uses the right interconnect for each link.

Marvell builds the silicon across the entire range of scale-out interconnect, so every link can be matched to the option that best fits the requirement.

The scale-out interconnect spectrum, from copper to co-packaged


Scale-out links fall on a continuum, from the shortest, lowest-power copper cables to the highest-performance retimed optics and co-packaged designs. Marvell silicon spans every class:

Interconnect typeWhat it isWhere it fits best
ACC
Active copper cable		Copper with a linear equalizer; no DSPThe shortest links: ~2m at 1.6T, between adjacent racks
AEC
Active electrical cable		Copper with a PAM4 DSP at each endShort rack-to-rack links: up to 7m at 1.6T
AOC
Active optical cable		Optical cable using the same PAM4 DSPs as optical modulesRack-to-rack optical beyond copper's reach: >30m at 1.6T
LPO
Linear-drive pluggable optics		Optical without DSP retimingShort optical links where power and latency matter most
TRO
Transmit-retimed optics		Retimes the transmit path onlyA balance of reach and power for use with switches equipped with high-performance SerDes
FRO
Full-retimed optics		Full DSP retiming, transmit and receiveThe longest, highest-bandwidth links across the floor
NPO
Near-package optics		Optical engine on the substrate near the packageField-replaceable with high density and power efficiency at the switch
CPO
Co-packaged optics		Optical engine inside the packageMaximum density and power efficiency at the switch

The same core technologies run across the entire spectrum. The Marvell portfolio of PAM4 DSPs — including the 5nm Nova series and the 3nm Ara series — powers retimed optical formats and AEC copper cables alike. Ara-series devices are tuned to each use case, including the Ara DSP for full-retimed optical modules and the Ara T DSP for transmit-retimed optical modules. Marvell TIAs and drivers are found in every optical format, and carry the signal path in LPO, where the DSP is removed, through a Marvell TIA and linear-driver chipset. Marvell optical engines are the foundation of co-packaged designs.

Whichever interconnect type an operator chooses, Marvell silicon can power it.

Will co-packaged optics take over the scale-out fabric?


Pluggable optical modules will lead scale-out connectivity for the next decade.

AI data center capex is accelerating, and the demand for fabric bandwidth is immediate. Pluggable optics are the only interconnect that can scale fast enough to meet it — backed by an ecosystem of 30+ module vendors that Marvell works with. Operators can deploy that capacity now, and shift volume from one vendor to another as any single supplier tightens — de-risking the build in a way a single-source approach cannot. Pluggables let the fabric be serviced, mixed, and upgraded link by link. That is why Marvell sees pluggables carrying the scale-out fabric for the next 8-10 years, with 1.6T modules ramping today and 3.2T on the way.

Co-packaged optics have a role. CPO offers the greatest density and power efficiency, right at the switch, and Marvell builds CPO silicon too, alongside every pluggable and cable format. Marvell supports every scale-out interconnect option so that it can address the specific needs of any given hyperscaler. But co-packaging trades away the serviceability, flexibility, and multi-vendor supply ecosystem that make pluggables the right answer for the immediate, massive build-out of scale-out fabric. Its natural home is the shortest, densest, most tightly co-designed connections: the scale-up domain.

Ethernet switching, purpose-built for the AI fabric


Interconnects carry the data; switching ties them into a fabric. These are the two pillars of connectivity. The switch at the center sets the power, latency, and shape of the entire cluster.

Distributed AI workloads rely on frequent, high-volume communication between compute nodes. Training large models and serving distributed inference pipelines generate sustained east-west traffic that places heavy demands on network infrastructure.

As clusters scale to hundreds of thousands or millions of accelerators, networks must support:

  • High aggregate bandwidth
  • Large fan-in and fan-out communication patterns
  • Predictable latency at scale
  • Power efficiency across thousands of links
  • Efficient congestion management and traffic coordination
Marvell® Teralynx® T100


The Marvell® Teralynx® T100 switch is designed to address these challenges. Its the industry's first at 102.4 Tbps that’s built specifically for the AI era. Where legacy switch architectures carry the overhead of the enterprise and cloud networks they were designed for, the T100 switch is a monolithic 102.4T device on 3nm that strips that overhead out — delivering up to 25% lower power than competitive solutions and the industry's lowest latency at this bandwidth tier, at under 1000W typical power consumption.

That efficiency is a scaling lever. As GPU and XPU racks approach 120kW, switching consumes an estimated 15–25% of rack power, so a lower-power switch lets operators place more accelerators inside the same power envelope. And with a 512-port radix, the T100 switch flattens the fabric: fewer tiers, fewer optical hops, lower latency and lower TCO across clusters of tens of thousands of accelerators.

  • Open by design. SONiC and SAI support and an OCP-aligned software toolkit, so the T100 switch drops into open, multi-vendor fabrics rather than a closed stack.
  • Standards-forward. Built for the latest Ultra Ethernet Consortium requirements and evolving AI Ethernet fabrics.
  • Proven lineage. The Teralynx data center switch portfolio spans 12.8T to 102.4T, a full range of switching for every tier of the data center.
The Marvell scale-out networking portfolio

The Marvell scale-out networking portfolio

One connected foundation, from copper between racks to optics across the floor, to the network switches that make the fabric:

Optical interconnect silicon

  • PAM4 optical DSPs. A multi-generational family, from the 5nm Nova DSP to the 3nm Ara series (the Ara, Ara T, and Ara X DSPs); first to 200G/lane; hundreds of millions shipped; 1.6T ramping today, with 3.2T demonstrated on 400G/lane electrical and optical signaling.
  • Coherent-lite optical DSPs. Used for long-reach scale out and campus scale-across applications, the Marvell Aquila coherent-lite DSP powers modules that combine the cost, power, and form factors needed for high-volume deployments with the performance-over-distance attributes of coherent modules.
  • TIAs and optical drivers needed for every optical module format and carrying the signal path in LPO.
  • Optical engines enabling near-package and co-packaged designs. 

Copper and optical cables

  • Alaska® A AEC DSPs and PAM4 DSPs for the active copper (AEC) and active optical (AOC) cables that wire adjacent and short rack-to-rack links.

Ethernet switching

Telemetry

  • The Marvell® RELIANT™ interconnect telemetry platform provides visibility and diagnostics across Marvell-powered interconnects, optical and copper alike, reducing operational complexity and improving reliability at scale.

Resources

Marvell to Showcase End-to-End AI Data Center Connectivity Portfolio at AI Infra Summit 2026

Marvell to Showcase End-to-End AI Data Center Connectivity Portfolio at AI Infra Summit 2026

Learn More

Marvell Wins Interconnect Product of the Year for Ara 3nm 1.6T PAM4 DSP

Marvell Wins Interconnect Product of the Year for Ara 3nm 1.6T PAM4 DSP

Learn More

The evolution of AI interconnects

The evolution of AI interconnects

Learn More

Scale-out networking FAQs

Scale out vs. scale up vs. scale across: how are they different? Arrow

Scale up tightly couples accelerators into a single compute domain — within a rack, and increasingly across a few racks. Scale out connects many such domains — row upon row of racks across a floor — into one cluster. Scale across connects separate data centers. Scale up makes many accelerators act as one domain, while scale out ties many domains into a cluster.

What is Ultra Ethernet, and why does it matter for scale-out networking? Arrow

Ultra Ethernet is the Ultra Ethernet Consortium (UEC) standard that adapts Ethernet for AI and HPC fabrics, adding the congestion management, telemetry, and transport features distributed AI training needs. It keeps scale-out open and multi-vendor while closing the performance gap with proprietary fabrics. The Marvell Teralynx T100 switch is built for the latest Ultra Ethernet Consortium standards.

Who supplies scale-out networking silicon for AI data centers? Arrow

Marvell builds the silicon across the whole scale-out fabric, both the interconnect and the switching. That spans PAM4 optical DSPs, coherent-lite DSP, TIAs, drivers, and optical engines across every module and cable format; AEC and AOC copper-and-optical cable silicon; Teralynx Ethernet switches built for Ultra Ethernet; and the RELIANT interconnect telemetry platform. 

What is scale out in AI infrastructure? Arrow

Scale out is the connectivity that links racks across many rows into a single AI cluster within a data center. It's the east-west fabric — often hundreds of thousands of links — that lets thousands or even hundreds off thousands of accelerators communicate and work as one during training and large-scale inference.

How is scale out different from scale up and scale across? Arrow

Scale up tightly couples accelerators into a single compute domain — within a rack, and increasingly across a few racks. Scale out connects many such domains — row upon row of racks across a floor — into one cluster. Scale across connects separate data centers. Scale up makes many accelerators act as one domain, while scale out ties many domains into a cluster.

What interconnect options does a scale-out fabric use? Arrow

A range, because the links aren't all the same. Copper cables (ACC, AEC) for the shortest rack-to-rack links; active optical cables (AOC) just beyond copper's reach; and pluggable optics in transmit-retimed (TRO), full-retimed (FRO) and linear (LPO) formats for longer links. Near-package optics (NPO) and co-packaged (CPO) optics offer maximum density at the switch. Marvell builds silicon across the entire range.

Will co-packaged optics replace pluggable modules in scale out? Arrow

The Marvell view is that pluggable optical modules will lead scale-out connectivity for the next decade. Pluggables are the only interconnect that can scale fast enough to meet immediate AI bandwidth demand, and they let operators service, mix, and upgrade the fabric link by link. Co-packaged optics have a role at the highest densities, and Marvell builds CPO silicon too, but its natural home is the shortest, densest links of scale up.

What role do high-radix switches play in scale-out networking? Arrow

High-radix switches reduce hop count and improve bandwidth utilization, simplifying network topology as clusters grow. 

Why is Ethernet used for scale-out fabrics? Arrow

Ethernet is open, multi-vendor, and scaling fast for AI through the Ultra Ethernet Consortium. It lets operators build large fabrics without locking into a single vendor. The Marvell Teralynx T100 switch brings 102.4T of purpose-built, low-power Ethernet switching to that fabric.

What does Marvell provide for scale out? Arrow

The full connectivity stack: PAM4 optical DSPs, coherent-lite DSPs, TIAs, drivers, and optical engines across every module and cable format; AEC and AOC cable silicon; Teralynx Ethernet switching; and the RELIANT interconnect telemetry platform matched to every link in the fabric.




Contact Us

We believe better partnerships help to build better technologies. Let’s connect and see what we can design together!

Thank You for Your Interest

We will be in touch with you soon!