Marvell Teralynx 10: Switch with 51.2 Tbit/s on one height unit

Marvell is launching a new switch on the market under the name Teralynx 10. It is designed in particular for the AI backend in data centers.

Save to Pocket listen Print view
Red and blue Ethernet cables hang from a rack

(Image: asharkyu/Shutterstock.com)

2 min. read
By
  • Benjamin Pfister

Marvell has announced the new Teralynx 10 switch. It is based on the open source network operating system SONiC (Software for Open Networking in the Cloud) from the Linux Foundation. The switch is aimed at operators of cloud data centers or AI clusters and offers up to 51.2 Tbit/s on one height unit and, according to Marvell, the lowest latency for data center switches.

Specifically, the switch supports either 64x 800 Gbit/s on two height units or 32x 1.6 Tbit/s on one height unit and can be used in various positions in the data center, such as leaf or spine. The manufacturer specifies a latency of less than 600 nanoseconds for forwarding in all packet sizes, which is particularly important for AI, ML and distributed workloads. Consumption is said to be 1 watt per 100 Gbit/s of bandwidth.

Thanks to a programmable forwarding architecture, the switch also enables new packet forwarding protocols without the need to redevelop the ASIC - a suitable addition to the open SONiC network operating system. The tunneling protocols currently supported include IP-in-IP, GRE, MPLS, VXLAN and Geneve, which is familiar from VMware NSX. Extended QoS and traffic management functions, including data center bridging (DCB) and RDMA over Converged Ethernet (RoCE), as well as the Precision Time Protocol (PTP) in accordance with IEEE 1588v2 are also integrated. The Teralynx Flashlight telemetry and analytics are designed to provide real-time visibility and granular network analysis to accelerate troubleshooting.

The data sheet for the Teralynx 10 can be found here. Marvell does not provide official pricing information.

(anw)

Don't miss any news – follow us on Facebook, LinkedIn or Mastodon.

This article was originally published in German. It was translated with technical assistance and editorially reviewed before publication.