While Teralynx T100 introduces several innovative features, let’s focus on three that fundamentally enhance the economics and performance of AI networking.

1. A Monolithic Design

Unlike other 102.4T switches, Teralynx T100 is designed around a single piece of silicon rather than a multi-die architecture. Multi-die designs lead to greater latency, bit error rates and power consumption because signals need to hop between the die in the package. And the more intense data traffic becomes, the more the benefits of a monolithic design get magnified.

Under normal workload conditions, the T100 will operate at a latency of approximately 400 nanoseconds—the lowest latency in this class, and even 20% lower than the 51.2T Marvell Teralynx 10 switch, the market-leader in latency for its class.³ Teralynx 100 will also deliver a deterministic latency—or an upper limit of latency over a sustained period of time—of under 420 ns.³ Lower latency directly reduces job completion time (JCT), an instrumental factor on TCO and ROI.

Teralynx T100 also uses up to 25% less power than competing solutions, consuming less than 1,000 watts under typical workload conditions thanks to its architecture and monolithic design. With these power savings, hyperscalers can thereby deploy more XPUs with the same power footprint.

Greater efficiency can also lower costs, making low-power switch silicon a strategic requirement. Switching and networking components consume approximately 15 to 25% of total rack power.⁴ Scale-out data center networking in the U.S. alone is expected to nearly double to 23 terawatt hours per year by 2028, according to Lawrence Berkeley National Laboratory.⁵ That’s as much power as EVs and plug-in hybrids combined consumed in the U.S. last year.⁶

2. Multiple Packaging and Connectivity Options

Customers now need to actively fine-tune fabrics for different use cases, applications and environments. As a result, Marvell is offering Teralynx T100 with different packaging solutions:

• Ball grid array (BGA) packaging solutions will be largely targeted at today’s predominant fabric architectures, i.e. scale-out networks built around high-speed switched and PAM4-based optical modules for inside-the-data-center connectivity and scale-across networks linking geographically distributed data centers with coherent ZR/ZR+ modules.
• Co-packaged copper (CPC) will become a leading option for creating efficient, economical scale-up networks that can fit within a single rack. CPC-based Teralynx switches can be combined with copper backplanes and flyover cables for a complete high-bandwidth copper solution that can meet loss and bit error rate requirements without additional retimers. CPC can also be combined with transmit-retimed optical modules, for efficient, mid-length links for scale-out networks.
• Co-packaged optics (CPO), meanwhile, will enable system architects to build multi-rack scale-up systems containing thousands of processors and terabytes of shared memory. High efficiency, high radix and low power consumption and latency will be absolutely critical here. CPO will also get deployed over time for scale-out networks.

In the next several years, connectivity will revolve around experimentation, with CPO, CPC and BGA-based all co-existing in data centers. By supporting the full spectrum, Teralynx T100 enables customers to optimize network design while simplifying validation and deployment.

3. Programmability for Real-world Performance Gains

Even the most advanced designs can’t anticipate every contingency. Teralynx T100 is fully programmable to enable customers to fine-tune networks and adapt to specific workloads, traffic patterns, and architectures. Programmable features include:

• Congestion Signaling (CSIG): Developed by Google and the Open Compute Project for traffic optimization, Teralynx T100 is the first switch to support CSIG for improved traffic management.
• Back-to-Sender (BTS): Enhanced fault detection for pinpointing problems and rerouting around them in real-time.
• Adaptive Routing: Introduced in T100, this feature facilitates the exchange of information across the fabric to make granular and real-time load balancing decisions on how to route traffic to relieve congestion.
• Flexible Pipelines and Microcode: Allows updates to support new protocols such as Ethernet scale up networking (ESUN) or the latest UEC requirements without a hardware upgrade. In addition, hardware resources such as TCAM and registers can be reallocated meaning Teralynx T100 can be reconfigured as a top-of-rack, spine, edge or AI cluster switch.

A New Era of Networking

The rapid innovation and the evolution of new use cases make today one of the most exciting times in decades to be part of networking. As AI workloads continue to grow, expect continued advances from Marvell like Teralynx T100 to meet these evolving demands.

1. Meta keynote OCP Summit, reprinted via Marvell, February 2022
2. 650 Group. Accelerated and Programmable Data Center Ethernet Switch and Quarterly Report and Market Forecast, March 2026
3. Marvell internal estimates
4. SemiAnalysis, June 2024
5. U.S. Data Center Energy Usage Report, Lawrence Berkeley National Library, 2024
6. U.S. Energy Information Administration, Electric Power Monthly, March 2026

# # #

This blog contains forward-looking statements within the meaning of the federal securities laws that involve risks and uncertainties. Forward-looking statements include, without limitation, any statement that may predict, forecast, indicate or imply future events or achievements. Actual events or results may differ materially from those contemplated in this blog. Forward-looking statements are only predictions and are subject to risks, uncertainties and assumptions that are difficult to predict, including those described in the “Risk Factors” section of our Annual Reports on Form 10-K, Quarterly Reports on Form 10-Q and other documents filed by us from time to time with the SEC. Forward-looking statements speak only as of the date they are made. Readers are cautioned not to put undue reliance on forward-looking statements, and no person assumes any obligation to update or revise any such forward-looking statements, whether as a result of new information, future events or otherwise.