Marvell has introduced the Teralynx T100, a 102.4 Tbps Ethernet switch chip aimed at AI and cloud data center networks. Marvell is calling it the industry’s first switch silicon at 102.4 Tbps “purpose-built for AI,” with a focus on reducing power and latency for large AI clusters.
The company puts typical power consumption for the Teralynx T100 at under 1,000 W. Marvell also claims up to 25% lower power than competing solutions at this bandwidth tier. For operators already fighting rack power limits, a sub-1 kW switch ASIC is a meaningful number—especially when the network fabric scales out into large switch counts.
Marvell says the T100 is a monolithic 102.4 Tbps device built on a 3 nm process, and that its design removes “legacy elements” intended for traditional enterprise and cloud data centers. The stated goal is to enable flatter, higher-radix fabrics by reducing the number of network tiers and optical links required in large AI deployments.
On the scale-out side, Marvell says the T100 supports up to a 512-port radix, targeting large AI training clusters with tens of thousands of accelerators. On the scale-up side, Marvell says its programmable pipeline supports multiple interconnect standards and emerging scale-up fabric protocols, including Ethernet Scale-Up Networking (ESUN), and aligns with the latest Ultra Ethernet Consortium (UEC) requirements for evolving AI Ethernet fabrics.
Those attributes map cleanly to the hard constraints showing up in modern AI builds: when systems push toward 120 kW per rack, every watt in the network matters, and so does latency added by extra tiers. But Marvell’s “lowest power” and “lowest latency” claims at 102.4 Tbps are the sort of figures that will hinge on real platform implementations, optics choices, and how the fabric is engineered end to end.
Marvell says the T100 is available in multiple package configurations, including ball grid array (BGA), co-packaged copper (CPC), and co-packaged optics (CPO), and includes integrated telemetry, congestion control, and traffic management logic for advanced data center architectures. The Teralynx portfolio also includes a software toolkit with an SDK, plus support for the Open Compute Project (OCP) Switch Abstraction Interface (SAI) and the SONiC network operating system.
“As AI workloads evolve and scale exponentially, hyperscalers require network architectures that optimize latency, power and scalability simultaneously,” said Rishi Chugh, vice president and general manager, Data Center Switch Business Unit, at Marvell.
Marvell expects the Teralynx T100 to begin sampling to customers this quarter.
Source: Marvell
