Marvell Technology has introduced the Teralynx T100, touted as the industry’s first switch silicon with an impressive capacity of 102.4 Tbps, specifically designed for artificial intelligence (AI) applications. This innovative silicon addresses the increasing demands of AI data centers, allowing for significant data throughput while maintaining low power consumption and low latency.
As AI workloads grow, data movement becomes crucial within data centers. Traditional clusters that once sufficed for basic applications are challenged by the sheer size and complexity of modern AI models, which require high-speed data handling across the entire infrastructure. This shift has heightened power demands, with GPU and XPU systems consuming up to 120KW per rack. Notably, networking components can account for 15-25% of the total rack power, making energy-efficient switching technology essential.
The Teralynx T100 offers up to 25% lower power consumption compared to competing solutions while achieving higher data rates. This advancement enables AI infrastructures to maximize the number of accelerators within the existing power frameworks, eliminating the need for further power supply resources.
Rishi Chugh, VP and general manager of the data center switch business unit at Marvell, emphasized that the Teralynx T100 is purpose-built for AI applications, designed to optimize performance without the constraints imposed by legacy technology. It aims to provide the reliability and efficiency needed for scaling future data center technologies.
For massive scale-out deployments, the switch can support up to a 512-port radix, facilitating network tier consolidation and latency reduction across extensive AI training environments equipped with thousands of accelerators. In scale-up scenarios, the device’s adaptable architecture accommodates various interconnect standards and cutting-edge protocols, including the Ethernet Scale-Up Networking (ESUN) protocol and evolving AI Ethernet fabrics.
The Marvell Teralynx T100 is slated to begin sampling to customers later this quarter.
For more on Marvell and their innovations: Marvell.
