Archive for the 'Networking' Category
-
May 21, 2026
PCIe-based Switching for AI Scale-up Networks
By Krishna Mallampati, Senior Director of Product Marketing, Data Center Switching, Marvell
Peripheral Component Interconnect Express (PCIe)® is the world’s most popular interconnect for connections between chips in a shared system, while ensuring low latency, and it is well suited to be deployed for the scale-up domain. Scale-up networks extend across racks and possess hundreds of processors; low latency and high bandwidth are required in these systems that make up the foundation of AI data centers.
Marvell demonstrates the industry’s first 260-lane PCIe 6.0 switch in the video below, marking a new performance standard for PCIe scale-up performance—256 lanes of data traffic (plus four lanes for management) is the industry’s highest radix for a PCIe switch.
Traditional PCIe switch architectures require multiple devices to scale, racking up complexity and cost. However, the Marvell Structera S PCIe switch flattens the network and eliminates the need for multiple smaller switches in a large scale-up system. This enables higher density, lower latency and overall increased system efficiency, making it an optimal solution for hyperscale operators.
-
May 19, 2026
Why Scale-up AI Networks Demand Scalable Optical Test
By Andrew Yick, Senior Director, Product & Test Engineering, Marvell
This article was first published in Photonic Integrated Circuits magazine.
The dominant challenge in modern AI infrastructure is not just the performance of a single accelerator but scaling up to thousands of accelerators (XPUs) in a cluster. Training and inference workloads now depend on an interconnect that can stitch these accelerators into a high-bandwidth, low-latency system, where performance is governed as much by the network as by the compute itself.
As these systems scale, physics asserts itself. Electrical links over copper hit a practical ceiling as routing density and channel loss collide, turning the loss bandwidth product into an impassable constraint. The choice is binary: either move electrical-to-optical conversion closer to the Application-Specific Integrated Circuit (ASIC) or surrender the link budget. Thus, to bypass this electrical wall, optics must migrate from the board edge and onto the ASIC package.
-
May 13, 2026
Solving AI’s Three Big Problems Through Photonic Fabric™ Technology
By Uday Poosarla, Senior Director, Product Management, Photonic Fabric Business Unit, Marvell
Three critical constraints are shaping the evolution of AI infrastructure:
- The scale-up network: Scale-up domains are increasing in size, processor count, and complexity. As these systems extend beyond a single rack, maintaining high bandwidth and low latency becomes significantly more challenging.
- The memory wall: There is a widening gap between the memory capacity and bandwidth AI workloads require and what existing memory subsystems can deliver efficiently.
- Performance per watt: Energy efficiency is becoming a defining constraint, both for controlling operating costs and staying within data center power limits.
Together, these challenges point to a common problem: the increasing difficulty of moving data efficiently across AI systems, from compute to memory across the scale-up domain.
The Marvell® Photonic Fabric™ technology platform addresses these challenges by combining the advantages of optical interconnect with system-level design innovation.
-
May 06, 2026
Scale-up Network Solutions for AI Infrastructure
By Preet Virk, Senior Vice President and General Manager, Photonic Fabric Business Unit
Modern AI infrastructure is built around multi-rack systems where thousands to tens of thousands of accelerators operate as a single logical compute element. As agentic AI and Mixture of Experts (MoE) models accelerate AI adoption, they are driving unprecedented scale and communication demands across data center infrastructure. These systems are connected by scale-up and scale-out networks that must deliver high bandwidth, low latency and efficient power. As these networks extend across racks, maintaining that performance becomes a primary challenge.
As AI systems grow in complexity and scale, the network becomes the backbone of the compute system. Large-scale clusters require massive XPU-to-XPU communication, driving an evolution beyond legacy protocols like PCIe® to encompass UALink™ (Ultra Accelerator Link), ESUN (Ethernet scale-up networking) and NVLink.
Meeting these requirements demands a new approach to connectivity. Marvell provides a comprehensive AI connectivity portfolio spanning scale-up, scale-out, scale-across and DCI (data center interconnect) network architectures. For scale-up networking, Marvell delivers copper and optical interconnects connecting XPUs, switches and memory. Within the rack, Marvell copper solutions provide low-latency, power-efficient short-reach connectivity, while Marvell optical interconnects enable high-performance scaling beyond the rack. This enables XPUs to operate as a more efficient, unified system as scale-up domains expand.
-
May 05, 2026
MACsec: A Shift in Security for Scale-across Networking
By Joseph Chon, Senior Director, Product Marketing, Data Center Interconnect, Marvell
MACsec is moving to the module in scale-across networks.
Media Access Control security (MACsec) is a foundational technology for protecting data in motion. It encrypts and authenticates Ethernet traffic to guard against eavesdropping, denial-of-service attacks, intrusion and other security threats while also strengthening overall data integrity. Embodied in silicon, MACsec further establishes a robust root of trust for managing encryption keys and securing the boot process.
What’s changing is where the silicon for delivering MACsec gets located.
To date, the MACsec circuitry for long-distance scale-across networks has typically been embedded in the switch ASIC, where space and silicon real estate are at an absolute premium. Embedding MACsec into the tight confines of the ASIC raises the cost of integrating the technology. It also makes infrastructure less flexible: some upgrades require taking the system offline, reducing overall capacity.
