Broadcom's StrataXGS Tomahawk 5 switch series provides 51.2 Terabits/sec of Ethernet switching capacity in a single monolithic device, which is twice the bandwidth of other available switch silicon.
“Delivering the world's first 51.2 Tbps switch, two years after we launched Tomahawk 4, the industry's first 25 Tbps switch, is a testament to the Broadcom team's outstanding execution and innovation,” said Ram Velaga, Vice -senior president and general manager of Core Switching Group. , Broadcom.
While data centers continue to experience dramatic growth in network bandwidth requirements, there is also a strong motivation to unify network infrastructure for general-purpose computing and storage with AI/ML computing. AI/ML training clusters are driving the need for meshes with high-bandwidth, high-baseline connectivity and shorter job completion times.
Ethernet offers the best solution for unified network infrastructure, providing lower power consumption, higher bandwidth, larger footprint and faster SerDes speeds, as well as a predictable doubling of bandwidth every 18 to 24 months. These benefits, combined with its large and vibrant ecosystem, Ethernet provides the highest performance interconnection per Watt and per dollar for AI/ML and cloud-scale infrastructure.
“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 more than a 95% reduction in power requirements,” added Velaga. “We applaud our customers, partners and engineers for making this possible.”
To enable the next generation of unified networks, Broadcom now offers the Tomahawk 5 family. Critical to enabling efficient use of widely shared infrastructure across large data centers, Tomahawk 5 provides AI/ML workload virtualization with features such as routing and single-pass VxLAN 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 bandwidth telemetry, and hardware-based link failover.
Tomahawk 5 Cognitive Routing improves network link utilization by automatically and dynamically selecting the links with the lowest system load for each flow traversing the switch. This is important for AI/ML workloads, which typically have a combination of short-lived mouse streams and long-lived, high-bandwidth elephant streams with low entropy.
Additionally, Tomahawk 5 includes real-time dynamic load balancing that tracks usage of all links, both on 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 diverts traffic from failed links. These features dramatically improve network utilization and reduce congestion, resulting in shorter JCT.
Also important to improving JCT is minimizing network congestion by controlling the rate of traffic injected into the network by each source. Because network operators employ a variety of different congestion control algorithms in their endpoints (such as commercial or custom NICs), Tomahawk 5 provides extensive programmable in-band telemetry on both live traffic and network probes.
Real-time metadata can be inserted into line-rate traffic as it traverses the network to collect telemetry about 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 consumption and lowest cost for physical connectivity, the Tomahawk 5 enables a PAM4 100G direct-to-direct-attach copper (DAC) interface, front-panel pluggable optics, and co-packaged optics. The flexible, long-range 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 with a broad ecosystem of standard pluggable optical modules on the front panel.
By leveraging Broadcom's cutting-edge silicon packaging and photonics technologies, the Tomahawk 5 will be available with co-packaged optics using Broadcom's Silicon Photonics Chiplets in Package (SCIP) platform, providing more than a 50% reduction in power required to optical connectivity. Because the same switch silicon offers all of these options, customers can choose the optimal I/O for each part of their intra-cluster, inter-cluster and inter-DC networks without the need for software portability.
Advantages of StrataXGS Tomahawk 5:
- Enables the next generation of unified data center infrastructure with 64 switching ports and 800GbE routing.
- Virtualization of general compute and AI/ML workloads with single-pass VxLAN routing and bridging.
- Unmatched physical I/O options using 512 instances of the industry's highest-performance, most flexible, and longest-range SerDes 100G PAM4.
- High-precision PTP and SyncE time synchronization.
- Six on-chip ARM processors for fully programmable, high-bandwidth streaming telemetry and sophisticated integrated applications such as on-chip statistics summarization.
- Unrivaled power efficiency, implemented as a 5nm monolithic 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 flow steering to Tomahawk 5. Hyperscale operators can now unify their network structures, eliminating specialized interconnects dedicated only to cluster training.”
Compared to general compute and storage, AI/ML training clusters have unique communication patterns. To minimize job completion time, Tomahawk 5 adds features specific to these workloads and network topologies.
StrataXGS Tomahawk 5 for AI/ML Features:
- The world's largest 200GbE port base: 256 ports supported on a single chip, enabling simple, low-latency AI/ML clusters.
- The industry's most advanced 51.2 Tbps shared buffer architecture, providing the highest performance and lowest end latency for RoCEv2 and other new RDMA protocols.
- Advanced Broadcom Cognitive Routing, dynamic load balancing, and support for end-to-end congestion control capabilities designed specifically to handle large, low-entropy streams typical of AI/ML workloads.
- Support for Clos and non-Clos topologies such as torus, Dragonfly, Dragonfly+, and Megafly.
- Hardware-based link failover to improve network resiliency and reduce JCT.
Along with the Trident and Jericho switch families, the Tomahawk series is part of Broadcom's three-pronged strategy of providing switch architectures optimized for different networking applications. All of 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 Broadcom SDK switch abstraction interface and 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 speed deployment time, support for SAI and Broadcom SDK is provided on Tomahawk 5 silicon, as well as a comprehensive set of network and device simulation tools.
