In what is claimed as the world’s first, Broadcom is now shipping its Tomahawk 6 switch series, capable of delivering 102.4 Terabits/sec of switching capacity in a single chip.
The switch platform is designed for modern AI workloads by doubling the bandwidth of any Ethernet switch available on the market built to power the next generation of scale-up or scale-out AI networks. Additionally, the company said it can meet the demands of AI clusters with more than one million XPUs.
According to the company, AI clusters are scaling from tens to thousands of accelerators that turns the network into a critical bottleneck. Breaking the 100 Tbps barrier allows these AI clusters to reach the next wave of AI infrastructure.
“[Tomahawk 6] marks a turning point in AI infrastructure design, combining the highest bandwidth, power efficiency, and adaptive routing features for scale-up and scale-out networks into one platform,” said Ram Velaga, senior VP and GM of the Core Switching Group at Broadcom. “Demand from customers and partners has been unprecedented. Tomahawk 6 is poised to make a rapid and dramatic impact on the deployment of large AI clusters.”
Features
Building on the CPO versions of Tomahawk 4 and 5, the Tomahawk 6 features 200G SerDes, which provides the longest reach for passive copper interconnect, enabling high-efficiency, low-latency system design with high reliability and low cost of ownership, Broadcom said.
The Tomahawk 6 will also be available with co-packaged optics while reducing link flaps and improving long-term reliability.
Broadcom said the Tomahawk 6 architecture allows for unified networks for AI training and inference. Cognitive Routing 2.0 in the switch features:
- Advanced telemetry
- Dynamic congestion control
- Rapid failure detection
- Packet trimming
These features allow global load balancing and adaptive flow control, Broadcom said.
