Feature articles |
By Nick Flaherty
Switching has become a key part of most networks, from the AI data centre down to the factory floor and now into the car.
While Broadcom and Marvell dominate the market for merchant Ethernet switch silicon in data centres through to industrial networks, NVIDIA has been producing switches based on three different protocols for its AI chips.
In the meantime, a new generation of data centre switch chips is using light rather than transistors, with a range of new technologies for higher performance and bringing new players into the market.
In the car, Microchip and NXP Semiconductors are driving switch silicon into vehicles to route many different types of data, from real-time control to infotainment.
NVIDIA’s switch technology covers a range of different protocols, from Ethernet to InfiniBand and a custom protocol called NVLink that is at the heart of its AI racks.
The sixth-generation NVLink is optimised for the Vera CPUs and Rubin GPUs to create racks of 72 GPUs connected by nine switch trays.
The sixth-generation NVLink enables 3.6 TB/s of bandwidth for each Rubin GPU, twice the bandwidth of the previous generation and more than 14x the bandwidth of PCIe Gen6. The NVLink 6 switch also introduces new management and resiliency features designed to maximise system uptime, including control-plane resilience, the ability to run with a partially populated rack, and hot-swapping of switch trays.
NVLink performance (Image: NVIDIA)
Other supercomputers use the InfiniBand protocol, so NVIDIA also has a series of switches for this. The Quantum-3 architecture used in the Q3400-RA and Q3401-RD switch chips enables 200Gbit/s-per-lane serializer/deserializer (SerDes) with 144 ports at 800Gb/s distributed across 72 octal small form-factor pluggable (OSFP) modules. The high connectivity, or radix, supports a two-level fat-tree topology capable of connecting up to 10,368 network interface cards (NICs) with minimal latency and optimal job locality, as well as other topologies providing connectivity to tens of thousands of GPUs.
All the Quantum-X800 switches include a dedicated OSFP InfiniBand in-band management port, which allows the full set of standard ports to be used for data network connectivity, simplifying port allocation and streamlining topology design.
The devices are adding silicon photonics with co-packaged optics to eliminate the need for pluggable optical transceivers to reduce electrical loss while enhancing signal integrity and improving overall power and thermal efficiency.
CPO technology reduces the high-speed electrical path to just a few millimetres within the substrate, slashing insertion loss to ~4 dB compared to 22 dB in traditional pluggable designs. This results in 63× better signal integrity, enabling higher data rates with lower DSP complexity and reduced power per bit.
The Spectrum-6 devices provide Ethernet switching optimised for NVIDIA’s full stack of network interface cards (NICs) and software at 104.6Tbit/s in AI racks. The DSX Air integrates simulation of Spectrum-X Ethernet with the full-stack AI factory, including components from NVIDIA and ecosystem partners. The optimisation allows full wire-speed throughput, ultra-low latency, and 160 MB per ASIC of fully shared packet buffer for fair and predictable performance.
Multiplane optimisation in the switch allows Ethernet to scale from thousands to hundreds of thousands of GPUs in a flat two-tier topology. Effective scaling requires overcoming challenges due to switch radix limits, oversubscription, and component reliability. The Spectrum-6 multiplane switch solves this with hardware to coordinate data routing. In a Spectrum-X multiplane network, each GPU connection is split into two or more ports, each of which connects into a separate network plane. Each network plane features its own distinct leaf/spine/optional layer called a superspine, and the NIC intelligently routes network traffic to different planes, load balancing packet by packet in hardware to give higher performance.
The Spectrum-6 switch chips are also adding CPO photonic silicon to integrate optical connections.
For the data centre, Broadcom’s merchant silicon networking architecture is organized into distinct families for applications. While hyperscale data centres and artificial intelligence (AI) fabrics use the high-throughput StrataXGS Tomahawk line, the modern Industrial Internet of Things (IIoT) and Industry 4.0 frameworks require different architectural profiles.
Industrial networking needs deterministic latency, deep buffering, multi-protocol programmability, and rugged environmental resilience. Broadcom’s switch roadmap transitions from the high bandwidth of Tomahawk to Trident for the network edge, the deep-buffered routing of StrataDNX Jericho, and ultimately the specialised, compact RoboSwitch family optimised for field-level deployment.
Tomahawk optimises for a monolithic, shared-buffer pipeline to achieve minimal port-to-port latency. The Tomahawk 4, used in the Stordis range of switches, achieves 25.6 Tbit/s with 50G PAM4 modulation for data centres, while the Tomahawk 5 uses 100G PAM4 SerDes to boost this to 51.2 Tbit/s, and Tomahawk 6 reaches 102.4 Tbps using ultra-dense 200G PAM4 signalling for AI networks.
Marvell Technology is known for its Prestera line of networking switch silicon for enterprise, edge, and carrier networks, but the acquisition of Innovium in 2021 brought the Teralynx platform for high-performance networks.
The Teralynx T100 is one of the first 102.4 Tbit/s switch silicon devices built on a leading-edge 3nm process. This provides 25% lower power consumption for AI racks, where the networking components can consume up to 25% of the energy.
The switches have a higher radix of up to 512 ports that allows operators to flatten network topologies. By reducing the number of switching tiers and optical links required to connect tens of thousands of AI accelerators, it drastically lowers both infrastructure costs and structural latency.
For AI training and inference, unpredictable latency can stall expensive GPU clusters. Marvell’s architecture ensures predictable, ultra-low latency under heavy, synchronized workloads.
The chip features a programmable forwarding pipeline that supports emerging standards like the Ultra Ethernet Consortium (UEC) and Ethernet Scale-Up Networking (ESUN) protocols.
Many data centre switch designs are structurally unsuited to the factory floor. They lack the extensive Layer 2/3 protocol tables and Time-Sensitive Networking (TSN) hooks required for deterministic hardware execution.
So Broadcom’s Trident introduces full user-plane programmability and extensive instrumentation for industrial aggregation and control-room environments.
Trident 3 and Trident 4 are powered by the programmable NPL (Network Programming Language) engine. These chips allow operators to deploy customised parsing pipelines via field upgrades. This is essential for handling encapsulated industrial payloads (such as EtherNet/IP, PROFINET, and Modbus TCP) alongside legacy IT traffic.
Trident 5-X12 brings advanced telemetry, MACsec line-rate encryption, and structural support for the IEEE Time-Sensitive Networking (TSN) standard suite.
For smart grids, mining operations, or distributed oil and gas fields, packet loss due to bursty traffic is unacceptable. So the Jericho switch uses an entirely different memory paradigm.
Unlike the on-chip SRAM architectures of Tomahawk, Jericho integrates high-bandwidth memory (HBM) directly onto the package substrate. This allows for massive, deep-packet Virtual Output Queue (VOQ) buffering that is up to 160 times larger than standard switch silicon. Jericho chips absorb massive microbursts caused by machine vision and high-frequency sensor spikes, ensuring zero packet loss. Jericho4 also implements high-performance Scale-Up Ethernet (SUE) and line-rate hardware encryption, making it ideal for the hardened industrial edge and distributed core routers.
At the field level, RoboSwitch targets low pin-counts, minimal power dissipation, and integrated Fast-Ethernet/Gigabit physical layers (PHYs).
The chip is designed for passive cooling and DIN-rail mounting with Single-Pair Ethernet (SPE / 10BASE-T1L), IEEE 1588v2 precision time stamping for synchronization, and hardware-based loop detection. This brings direct Ethernet connectivity down to the isolated sensor, valve, and actuator level under extreme thermal ranges in industrial applications.
As vehicles are more complex with dramatically higher levels of compute power, switching becomes more important.
The Microchip LAN9371 is a scalable, compact, and cost-effective multi-port AVB/TSN 100BASE-T1 Ethernet switch based on the IEEE 802.3bw-2015 specification. The LAN9371 incorporates a Layer-2+ managed high-performance Ethernet switch, three 100BASE-T1 physical-layer transceivers (PHYs), and two MAC ports with individually configurable RGMII/MII/RMII interfaces for direct connection to a host processor/controller, another Ethernet switch, or an Ethernet PHY transceiver. An additional IEEE 802.3/802.3u-compliant 100BASE-TX port is provided for applications where an integrated automotive OBD port is required.
The LAN9371 is available in a Grade 2 Automotive (-40°C to +105°C) temperature range and is qualified to AEC-Q100 for gateways, Advanced Driver Assistance Systems (ADAS), infotainment, telematics, and in-vehicle networking. The LAN9371 fully supports the IEEE family of Audio Video Bridging (AVB) standards, which provide high Quality of Service (QoS) for latency-sensitive traffic streams over Ethernet. Hardware time-stamping and time-keeping features support IEEE 802.1AS (gPTP) and IEEE 1588v2 (PTP) time synchronisation.
All ports feature eight egress queues and an IEEE 802.1Qav credit-based traffic shaper and time-aware scheduler, as per the IEEE 802.1Qbv specification. A host processor can access all LAN9371 registers for control over all PHY, MAC, and switch functions.
NXP Semiconductors has shifted to Ethernet switches for automotive designs, ranging from 5-port 100Mbit/s devices up to advanced 10-port multi-gigabit engines that support Layer 2/3 routing for a resilient vehicular backbone.
The SJA1105 and the highly integrated SJA1110 and SJA1610 multi-gigabit families are specifically engineered to handle the high-bandwidth, ultra-low-latency data pipelines required for Advanced Driver Assistance Systems (ADAS), autonomous driving vectors, and connected infotainment hubs.
The devices natively implement the IEEE TSN standard to guarantee that safety-critical control traffic such as braking commands takes absolute priority over non-critical data. Features like IEEE 802.1Qbv for time-aware shaping and IEEE 802.1Qbu/802.3br for frame preemption allow the switch to interrupt a large, low-priority infotainment packet mid-transmission to let a critical control frame pass, ensuring bounded, sub-microsecond latency.
The switches incorporate hardware-enforced Deep Packet Inspection (DPI) and Access Control Lists (ACLs). This allows the switch to inspect incoming frames at line rate, dropping malformed packets or unauthorized intrusion attempts before they can reach the vehicle’s primary Electronic Control Units (ECUs). They also include secure boot capabilities to ensure firmware integrity.
The next generation of network switches is using light natively to avoid the penalty of converting to and from electricity. This is opening up several new suppliers, such as Coherent, Lumentum, Axonal and even connector giants Huber + Suhner and Molex.
Optical circuit switching (OCS) has become one of the fastest-growing segments in networking, with revenue expected to exceed $3.5 billion by 2029, more than twice that of 2025.
The OCS developed by Coherent optimises data centre networks by minimising electrical switches and optical-electrical-optical (OEO) conversions using liquid crystal. This can result in significant cost savings, reduced power consumption, and improved latency for GPU connections.
The switch is available in a range of configurations, including 64×64 to 320×320 and 512×512 for AI front-end and back-end interconnect fabrics.
Lumentum is working with Marvell to integrate an OCS with digital signal processing (DSP) modules for 1.6Tbit/s data rates. Lumentum uses MEMS micromachined silicon mirrors to switch the light, handling 300 x 300 ports for large AI systems and 64 x 64 for smaller systems.
The MEMS mirrors have a lifetime of 1 trillion operations to create low-latency direct optical paths that are 10× lower than packet‑switched alternatives. The switching is protocol‑agnostic and supports Ethernet, InfiniBand and proprietary links at any data rate within the O and C wavelength bands. A proprietary MEMS control system eliminates closed‑loop dither and the fully passive optical switching core and modular design simplifies service and minimizes downtime.
Marvell used its new RELIANT software platform to analyse equipment performance and optimise the network in real time, as well as monitor data transmission, power consumption, bit error rate, and other metrics.
Molex is also using MEMS mirrors for its High-Radix OCS Platform. This supports up to 544×544 ports with a roadmap to support 1,000+ ports, enabling fewer switch tiers and hops with software-reconfigurable connectivity as workloads evolve.
Low insertion loss extends optical reach with fewer amplifiers and regenerators, while sixteen hot-swappable MEMS driver cards help limit impact during on-site field service, reducing downtime. An OCS Network Operating System helps operators integrate management tools and automatically recalibrate mirrors.
Meanwhile, Huber + Suhner has an OCS called Polatis that uses piezoelectric actuators to steer the optical beams into the output fibres with minimal loss, distortion, or interference between paths. Switches range from 8 x 8 to 384 x 394, and alignment is maintained using feedback from integrated position sensors to ensure connection stability over time, temperature, and external disturbances.
This technology, called DirectLight, means that the Polatis switching occurs completely independently of the power level, colour or direction of light on the path. This is different from other native optical switches and enables the switch to hold optical connections on unlit paths, which permits the pre-provisioning of dark fibres and allows bidirectional transmission.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。