Marvell Technology announced the production and customer deployment of its flagship 51.2 Tbps programmable Teralynx 10 Ethernet switch chip for cloud data centers. The Teralynx 10 is engineered to support training, inference, general-purpose computing, and other workloads, driving the scale of accelerated infrastructure in the rapidly growing AI sector.
Marvell said its Teralynx 10 is designed with a clean-sheet switch architecture, delivering high bandwidth, ultra-low latency, low power consumption, high radix, and zero-latency-loss programmability. This unified platform can be deployed across various network points without sacrificing functionality or performance. The switching silicon could be used in top-of-rack (ToR), leaf, spine, AI clusters, and network edge applications. Marvell’s Teralynx 10 also supports the industry’s transition to open networking, offering a robust solution that ensures a broad ecosystem of switch manufacturers and enhances silicon diversity and supply chain stability for cloud network operators.
Marvell also announced that its Teralynx 10 is available within the Linux Foundation’s SONiC (Software for Open Networking in the Cloud) network operating system. Teralynx 10-based switches can be used to power open-network environments, enabling cloud data center operators to tailor their systems, accelerate development, and diversify silicon vendors.
Key Specifications:
- Capacity: 51.2 terabit-per-second Ethernet throughput with large buffers. Integrates Marvell’s proven, robust 112G LR SerDes, supporting up to 64 ports of 800GbE or 128 ports of 400GbE.
- Latency: As low as 500 nanoseconds, with sub-600 nanoseconds latency across all packet sizes with cut-through & store-and-forward modes
- Radix: 512 switching radix, reducing the number of switch tiers in large clusters. OEM & ODM switches can provide high port count 100G, 200G, 400G and
800G connectivity using 50G/100G PAM4 with support for 25/50/100/
200/400 and 800 GbE connectivity - Power Consumption: 1 watt per 100 gigabits-per-second of bandwidth.
- Programmability: Fully programmable architecture with no impact on packet processing capacity or latency.
- Traffic Management: Advanced QoS/traffic management feature set including DCB and RoCE. Highly scalable/flexible layer 2 and 3 tables for IPv4, IPv6 and hybrid networks. Tunneling protocols including IP-in-IP, GRE, MPLS, VXLAN and Geneve.
- Advanced Teralynx Flashlight telemetry & analytics: A ground-up telemetry architecture that delivers extensive real-time visibility and actionable granular network analytics to troubleshoot and resolve network issues quickly
- Ecosystem: Supported by major OEMs, ODMs, and ISVs to facilitate adoption and optimization.
Marvell notes that its cloud-optimized Teralynx 10 and Nova networking platform ensures interoperability between switch and optics, thereby reducing the burden of validation and interoperability testing for customers and accelerating the deployment of these next-generation technologies.
The introduction of the Teralynx 10 comes at a pivotal time as the demand for high-bandwidth connectivity in AI data centers surges. This switch device supports the shift from proprietary network operating systems (NOS) to open network platforms such as the Linux Foundation’s SONiC and SAI, facilitating faster deployment and optimization across multiple manufacturers. According to 650 Group, shipments of 51.2 Tbps switches are projected to rise dramatically from 77,000 units in 2024 to 1.8 million by 2028, reflecting a 120% compound annual growth rate (CAGR). This significant growth underscores the switch’s critical role in the next generation of data center infrastructure.
Industry partners cited in the announcement include Celestica, Wistron Neweb, Keysight, MultiLane, and Teledyne LeCroy Xena.
“AI deployments require a switch solution which is simultaneously high bandwidth, low power, low latency, and future proofed for evolving requirements. Marvell has delivered the most complete AI Ethernet switch solution available to the industry today, and we are pleased to deliver a production-ready solution for customers' expanding 51.2 Tbps deployments,” said Nick Kucharewski, senior vice president and GM, Network Switching Business Unit, at Marvell. “SONiC has emerged as the clear solution to enable open switch platform interoperability as well as silicon vendor diversity for the world’s largest cloud hyperscalers.”
