Broadcom has launched its latest data center networking silicon, the Tomahawk 6, which boasts a remarkable bandwidth of 102.4 Terabits per second (Tbps), doubling the performance of its predecessor, the Tomahawk 5, which delivered 51.2 Tbps. This advancement is crucial for hyperscale operators preparing to deploy extensive clusters with up to one million XPUs (processing units such as GPUs, TPUs, and other AI accelerators).
Pete Del Vecchio, a product manager for the Tomahawk switch family, emphasizes that this leap in raw bandwidth signifies the most substantial performance increase in the Tomahawk family to date. However, achieving this enhancement entails significant engineering complexities beyond mere linear scaling. For instance, as bandwidth doubles, on-chip entities like the memory management unit (MMU) become approximately four times more complex, posing considerable challenges in packet buffering, queue scheduling, and congestion control.
The Tomahawk 6 introduces several architectural innovations tailored to meet the needs of modern AI workloads. It supports configurations with either up to 1,024 100G SerDes lanes or alternative high-speed 200G SerDes options, allowing flexibility for various deployment scenarios. Moreover, the chip’s capability to support longer passive copper interconnects provides a cost-effective solution compared to optical alternatives while reducing power consumption.
One of the primary features of Tomahawk 6 is its capacity to manage both scale-up and scale-out networking requirements within a cohesive Ethernet framework. This dual capability streamlines network operations by eliminating the necessity for distinct technologies and protocols for these different tiers of networking.
Additionally, Tomahawk 6 leverages Cognitive Routing 2.0, an adaptive routing technology designed specifically for AI workloads. This upgraded system integrates advanced telemetry, dynamic congestion control, and rapid failure detection to maintain performance across large-scale training jobs, addressing congestion hotspots effectively.
The development of the Tomahawk 6 was a significant undertaking, requiring over three years of intensive engineering work and a substantial investment in architecture, silicon design, and validation processes. Broadcom asserts that it holds a competitive edge in switching bandwidth, with no rival Ethernet switch chip reaching the capabilities of Tomahawk 6, while also maintaining an advantageous time-to-market position.
Broadcom’s strategic focus on meeting the growing demands of AI networking positions the Tomahawk 6 as a critical player in the surge of AI and large-scale computing workloads.
