Broadcom's 102.4 Tbps Tomahawk 6 targets million-XPU AI clusters

The MMU complexity increase reflects the challenges of managing packet buffering, queue scheduling and congestion control at these extreme bandwidths. Traditional approaches to packet switching become increasingly difficult as the numbers of ports, queues and simultaneous flows grow exponentially.

The Tomahawk 6 addresses these challenges through several key architectural innovations. The chip supports configurations with up to 1,024 100G SerDes lanes or higher-speed 200G SerDes options, providing flexibility for different deployment scenarios. For AI clusters requiring extended reach, the 100G SerDes configuration enables longer passive copper interconnects, reducing both power consumption and total cost of ownership compared to optical solutions. (Read more: Copper-to-optics technology eyed for next-gen AI networking gear)

Unified scale-up and scale-out architecture

One of Tomahawk 6’s most significant technical achievements is its ability to handle both scale-up and scale-out networking requirements within a unified Ethernet framework. 

Scale-up networking refers to high-bandwidth, low-latency connections within individual AI training pods, typically supporting up to 512 XPUs in the Tomahawk 6’s case. Scale-out networking connects these pods together into larger clusters, with Tomahawk 6 supporting deployments exceeding 100,000 XPUs.

This unified approach eliminates the need for separate networking technologies and protocols between scale-up and scale-out tiers, simplifying network operations. 

AI-optimized routing and congestion control

The Tomahawk 6 incorporates Cognitive Routing 2.0, an enhanced version of Broadcom’s adaptive routing technology specifically designed for AI workloads. This system provides advanced telemetry, dynamic congestion control, rapid failure detection and packet trimming capabilities that enable global load balancing across the network fabric.