Tomahawk 5: Driving the Future of AI/ML Networking

The new Broadcom StrataXGS® Tomahawk 5® is a triumph of engineering innovation that paves the way for a new era of cloud computing— affirming the spirit of Connected by Broadcom philosophy. It enables the more powerful, efficient, environmentally responsible, and affordable data centers needed for the compute, storage, and AI/ML processing in future hyperscale cloud infrastructure. Tomahawk 5 is the world’s first and only 51 Terabit per second (Tbps) switch, providing double the bandwidth of any other switch available today.

This new silicon chip continues the industry-leading legacy of Broadcom innovation in the Ethernet switching space, a record that includes doubling the bandwidth of its Tomahawk series every two years since its first version was introduced in 2014. To put this history in perspective, a single Tomahawk 5 replaces forty-eight Tomahawk 1 switches in a data center network, resulting in a 95 percent reduction in energy consumption.

So, how did we accomplish this and what does it mean?

Turning Moore’s Law on Its Head

The basic hypothesis of Gordon Moore’s famous theorem, that the number of components in an integrated circuit doubles every two years, is now reaching its physical limits. It is becoming impossible to continue scaling chips further and still conform to the exponential growth predictions of Moore’s Law.

The sunsetting of Moore’s Law creates challenges in planning and developing new cloud-scale networking infrastructure. The sheer size of these hyperscale networks and their concurrent need for bigger data centers to house them, compounded by the tremendous bandwidth requirements of new AI/ML workloads, is increasingly problematic.

There is growing consensus that what’s needed is a solution that can unify the various components of data center infrastructure, from general-purpose compute to storage to highly specialized and more bandwidth intensive AI/ML accelerators. A solution is needed that reduces infrastructure size, power requirements, environmental impact, and overall cost.

Tomahawk 5 answers that need by basically turning Moore’s Law on its head. It reinvents the integrated circuit to do far more with far less.

Tomahawk 5 is specifically designed to meet the needs of massive data centers by eliminating the barriers to more efficient and faster processing. Rather than building more complexity and processing responsibilities into the chip, it fundamentally does the opposite.

Tomahawk 5 is part of Broadcom’s three-prong strategy to Ethernet switches, where domain-specific switch chips are created to focus on the needs of different markets.  Combined with the Trident and Jericho families, which focus on the Enterprise and Service Provider markets, respectively, Tomahawk is fully optimized for the needs of hyperscale data centers, reducing chip functionality to only the features needed for hyperscale data centers.

Tomahawk 5 leverages the advantages of an open Ethernet ecosystem, AI/ML technologies, and other software innovations to enable the highest performance networks with the lowest cost and lowest energy consumption. In a sense, Tomahawk 5 essentially assimilates a large portion of the data center’s network on a single chip and, echoing a phrase that many might recall from the glory days of Sun Microsystems, makes the network the computer.

The chip’s monolithic design eliminates chip-to-chip interfaces, enabling efficient data communication within the chip. This high level of integration vastly improves the network’s speed, enabling dramatic performance improvements for the most exciting next-generation technologies including virtual reality, immersive scale video experiences, driverless cars, and other forms of autonomous vehicle transportation. As a single monolithic chip, it also provides the highest energy efficiency.

One Chip to Unify Them All

Tomahawk 5 also parallels the remarkably swift advance of AI/ML technology.  Features built into Tomahawk 5 handle many of the most critical functions driving scale-out networking, including all the traffic management functions critical to enabling better and more efficient use of the massively shared infrastructure in large data centers – a huge advantage over proprietary network fabric solutions.

Tomahawk 5 automatically selects the most efficient way to transfer information across the network by tracking utilization of all links both at the switch itself and downstream. It monitors the loading and health of these links and steers traffic away from failed links.  The result is dramatically improved network utilization, reduced congestion, and shortened job completion time (JCT) for even the largest AI/ML workloads.

Tomahawk 5 also demonstrates why Ethernet, with advantages that include the lowest power consumption, highest bandwidth, and scalability to the highest number of compute nodes, is the best solution for creating a unified network infrastructure at hyperscale. The benefits include:

• Enables next generation unified data center infrastructure
• Extends virtualization across all data center workloads
• Provides highest performance I/O options
• Delivers unmatched power efficiency

The foundation for the next generation of cloud scale computing and AI/ML infrastructure is here. Drive that future today with Tomahawk 5, Connected by Broadcom.

Learn More Now:

• Read more the new Broadcom StrataXGS® Tomahawk 5®
• Read the Broadcom StrataXGS® Tomahawk 5® PR Launch Announcement