Broadcom Inc. announced that it has delivered the StrataXGS® Tomahawk® 5 switch series, providing 51.2 Terabits/sec of Ethernet switching capacity in a single, monolithic device, double the bandwidth of any other switch silicon available on the market today.
“Delivering the world’s first 51.2 Tbps switch two years after we released Tomahawk 4, the industry’s first 25 Tbps switch, is a testament to the outstanding execution and innovation by the Broadcom team,” said Ram Velaga, senior vice president and general manager, Core Switching Group, Broadcom. “Since the introduction of Tomahawk 1 in 2014, Broadcom has consistently executed on doubling the bandwidth approximately every two years. With today’s introduction of the fifth generation Tomahawk family, we are proud to say that a single Tomahawk 5 replaces forty-eight Tomahawk 1 switches in the network, resulting in over 95 percent reduction in power requirements. We applaud our customers, partners, and engineers for making this possible.”
While data centers continue to experience dramatic growth in network bandwidth requirements, there is also strong motivation to unify the networking infrastructure for general-purpose compute and storage with that of AI/ML compute. AI/ML training clusters are driving the need for fabrics with high-bandwidth connectivity, high radix, and lower job completion time, while operating at high network utilization.
Ethernet offers the best solution for unified network infrastructure, providing the lowest power, highest bandwidth, highest radix, and fastest SerDes speeds, along with a predictable doubling of bandwidth every 18 to 24 months. These benefits combined with its large and vibrant ecosystem provide the highest performance interconnect per Watt and per dollar for AI/ML and cloud scale infrastructure.
To enable the next generation of unified networks, Broadcom is now offering the Tomahawk 5 family. Critical to enabling efficient use of the massively shared infrastructure in large data centers, Tomahawk 5 provides AI/ML workload virtualization with features such as single-pass VxLAN routing and bridging. Critical to minimizing job completion time (JCT) for AI/ML workloads, Tomahawk 5 offers features such as Broadcom Cognitive Routing, advanced shared packet buffering, programmable inband telemetry, and hardware-based link failover.
Tomahawk 5’s cognitive routing improves network link utilization by automatically and dynamically selecting the most lightly loaded links in the system for every flow that transverses the switch. This is especially important for AI/ML workloads which often have a combination of short-lived mice flows and long-lived, high-bandwidth elephant flows with low entropy. Tomahawk 5 includes real-time dynamic load balancing that tracks the utilization of all links, both at the switch, and downstream in the network, to determine the optimal path for each flow. It also monitors the health of links in hardware and automatically steers traffic away from failed links. These features provide dramatically improved network utilization and reduced congestion, resulting in shortened JCT.
Also important in improving JCT is minimizing network congestion by controlling the rate of traffic injected into the network by each source. Since network operators employ a variety of different congestion control algorithms at their endpoints (such as merchant or custom NICs), Tomahawk 5 provides extensive programmable inband telemetry on both live traffic and network probes. Real-time metadata can be inserted into traffic at line rate as it traverses the network to collect telemetry on queue size, packet latency, switch utilization, and a variety of other customer-selectable metrics. This metadata can be used for precise end-to-end network congestion control.
To enable the lowest power and lowest cost for physical connectivity, Tomahawk 5 enables a direct 100G PAM4 interface to direct attach copper (DAC), front panel pluggable optics, and co-packaged optics. The flexible, long reach Tomahawk 5 SerDes provides DAC connectivity to all devices within a rack, and even between racks, without the need for retimers or other active components. It can also interface directly to a broad ecosystem of standard front panel pluggable optical modules.
Additionally, leveraging Broadcom’s leading-edge silicon photonics and packaging technologies, Tomahawk 5 will be made available with co-packaged optics using Broadcom’s Silicon Photonics Chiplets in Package (SCIP) platform, providing more than 50 percent decrease in the power needed for optical connectivity. Since the same switch silicon provides all these options, customers can choose the optimal I/O for each part of their intra-cluster, inter-cluster, and inter-DC networks with no software porting required.
StrataXGS Tomahawk 5 Series Key Benefits:
- Enables the next generation of unified data center infrastructure with 64 ports of 800GbE switching and routing.
- Virtualization of general compute and AI/ML workloads with single-pass VxLAN routing and bridging.
- Unparalleled physical I/O options using 512 instances of the industry’s highest performance, most flexible, and longest-reach 100G PAM4 SerDes.
- High-precision PTP and SyncE time synchronization.
- Six on-chip ARM processors for high-bandwidth, fully-programmable streaming telemetry, and sophisticated embedded applications such as on-chip statistics summarization.
- Unmatched power efficiency, implemented as a monolithic 5nm die.
“Tail latency is the critical network performance metric for distributed AI/ML training,” said Bob Wheeler, principal analyst at Wheeler’s Network. “Broadcom recognized the limitations of traditional hash-based load balancing for these workloads and added cognitive routing with dynamic flowlet steering to Tomahawk 5. Hyperscale operators can now unify their network fabrics, eliminating specialized interconnects dedicated to only training clusters.”
Compared to general compute and storage, AI/ML training clusters have unique communication patterns. To minimize JCT, Tomahawk 5 adds specific features for these workloads and network topologies.
StrataXGS Tomahawk 5 Series Key Features for AI/ML:
- World’s highest radix of 200GbE ports: 256 ports supported on a single chip, enabling flat, low latency AI/ML clusters.
- The industry’s most advanced 51.2 Tbps shared-buffer architecture, providing the highest performance and lowest tail latency for RoCEv2 and other RDMA protocols.
- Advanced Broadcom Cognitive Routing, dynamic load balancing, and support for end-to-end congestion control capabilities specifically designed to handle the large, low entropy flows typical of AI/ML workloads.
- Support for Clos and non-Clos topologies such as torus, Dragonfly, Dragonfly+, and Megafly.
- Hardware-based link failover for improved network resiliency and reduced JCT.
Along with the Trident and Jericho switch families, the Tomahawk series is part of Broadcom’s three-pronged strategy of providing optimized switch architectures for different network applications. All these devices share a common programming interface, so customers can easily leverage their software development efforts across different platforms.
Having a strong commitment to open networking, Broadcom has provided both Switch Abstraction Interface (SAI) and Broadcom SDK open APIs for all five generations of the Tomahawk family. Broadcom is one of the industry’s largest contributors to SAI and the SONiC network operating system. To accelerate time to deployment, support for SAI and Broadcom SDK are provided on Tomahawk 5 silicon as well as a comprehensive suite of network and device simulation tools.
Broadcom Tomahawk 5, BCM78900, is available today for customers worldwide. For more information on Tomahawk 5 click here.