Unique architecture delivers high bandwidth, low latency, low power and programmability in a single platform for optimized cloud AI deployments.
Marvell Technology, Inc., a leader in data infrastructure semiconductor solutions, announced that the Marvell® Teralynx® 10 Ethernet switch device is in volume production with customer deployments underway. The Teralynx 10 switch is a low power, programmable 51.2 Tbps Ethernet device with the industry’s lowest latency1, delivering leading performance for training, inference, general-purpose compute and other workloads to scale accelerated infrastructure in cloud data centers.
The Teralynx 10 switch device enters production as the growth in the size of AI clusters and the number of AI data centers is driving increased demand for high-bandwidth connectivity. Simultaneously, the industry is undergoing a shift from the exclusive use of proprietary network operating systems (NOS) to deployment of open network platforms, such as Linux Foundation’s SONiC and SAI (switch abstraction interface), that enable operators to rapidly deploy switching from multiple manufacturers, speed up development timelines and optimize for their specific use cases.
Shipments of 51.2 Tbps switches are expected to skyrocket from approximately 77,000 in 2024 to 1.8 million in 2028, a 120% compound annual growth rate (CAGR), according to 650 Group. 650 Group’s co-founder Alan Weckel said, “51.2T marks a watershed moment in data center switching. Cloud service providers will rapidly gravitate toward this switch capacity to build the next generation of infrastructure, placing a priority on multi-vendor diversity at the system and silicon levels.”
The Teralynx 10 device is based on a clean-sheet switch architecture that simultaneously delivers high bandwidth, ultra-low latency, low power, high radix and zero-latency-loss programmability. This enables a single switching platform design to be used across multiple points of the network without compromising functionality or performance. Whether used in top-of-rack (ToR), leaf, spine, AI cluster, or network edge, the unified platform offers the attributes required for advanced AI networks and enables in-field feature updates to support new protocols.
The Teralynx 10 switch software platform is designed to deliver on the emerging transition to open networking for large scale data centers. Increasingly, operators are moving from traditional proprietary network operating systems to the open-source SONiC platform, similar to how Linux disrupted proprietary operating systems for server compute. An open-source approach enables operators to normalize their networking system software across multiple manufacturers and accelerate development timelines. From a business perspective, this approach allows for a broader ecosystem of switch manufacturers with increased silicon diversity and supply chain stability for cloud network operators. Operators are increasingly mandating open network software, and the Marvell Teralynx 10 switch platform provides a comprehensive solution with an industry-accepted programming model that can run on Marvell switch solutions as well as other compatible hardware. (See today’s related announcement on SONiC.)
Also Read: Brenmiller to Expand bGen™ Capabilities for AI Data Center Cooling Applications
“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. “This is the right product at the right time, as the market turns to innovative solutions from customer-focused suppliers like Marvell to satisfy their explosive demand for high bandwidth switching in AI data center buildouts.”
Teralynx 10 Features and Benefits
The Teralynx 10 Ethernet switch offers leadership performance and high scale for the most advanced cloud and AI workloads:
- Lowest latency: Demonstrates 51.2 terabit-per-second Ethernet throughput with latency as low as 500ns, and sub-600 nanoseconds latency across all packet sizes. Low latency is essential for meeting the demands of AI, ML and distributed workloads, and it directly impacts job completion time and algorithmic efficiency. The low latency Teralynx switch gives cloud operators the ability to reduce operating expenses and increase their capacity for performing revenue-generating activity.
- Industry-best radix: 512 switching radix enables operators to reduce the number of switch tiers in large clusters, yielding dramatically lower power and total cost of ownership (TCO).
- Low power consumption: The switch consumes 1 watt per 100 gigabits-per-second of bandwidth.
- Programmable: The industry’s only switch architecture that is fully programmable with no impact to packet processing capacity or latency. The Teralynx 10 device can be applied in multiple use cases at 51.2 Tbps. This flexibility gives data center operators investment protection as networking technologies evolve to handle new protocols.
Expansive Partner Ecosystem
Switch systems based on the Teralynx 10 device are available from major OEMs, ODMs, and ISVs to speed adoption, while providing greater choice and ability for customers to optimize for their unique environment.
“The proliferation of AI and machine learning in data centers represents a transformative shift in the industry, driving unprecedented advancements in computational power, data processing capabilities, and operational efficiency,” said Gavin Cato, head of Portfolio Solutions and CTO, Connectivity and Cloud Solutions, Celestica. “Celestica congratulates Marvell on the Teralynx 10 51.2T release and looks forward to continued collaborations in leading-edge Ethernet solutions to support bringing advanced AI capabilities to the market.”
“We have been collaborating with Marvell to provide highly innovative and comprehensive solutions for our customers and we are very excited to add Teralynx 10 to our portfolio,” said David Tsai, vice president and GM, Wistron Neweb Corporation, Networking Business Group.
“Keysight’s industry-leading test solutions, like the AresONE-M 800GE platform, were instrumental in accurately benchmarking the Teralynx 10 low-latency capabilities. This precise measurement of data transfer through the switching fabric is critical in assessing optimal job completion times for complex AI training algorithms within massive data center clusters exceeding 100,000 GPUs. By collaborating with Marvell, we’re empowering customers to confidently deploy the Teralynx 10 and achieve peak performance in their AI cloud infrastructure,” said Ram Periakaruppan, vice president and general manager, Network Test & Security Solutions, Keysight.
“MultiLane is proud to extend its partnership with Marvell by introducing the Teralynx 10 platform into our Interoperability Labs in Fremont, Calif. As a leader in interoperability and compliance testing services, MultiLane continues to enhance its offerings to help data center operators streamline the adoption and deployment of high-performance switching fabrics,” said Michel Haddad, general manager, Measurement Solutions Business Unit, MultiLane.
Source: PRNewswire
