Broadcom's Tomahawk Ultra: Revolutionizing Ethernet for AI and HPC

Broadcom's Tomahawk Ultra: A Game-Changer for AI and HPC Networking

Broadcom has announced the shipment of its groundbreaking Ethernet switch, the Tomahawk Ultra, designed to revolutionize high-performance computing (HPC) and AI workloads. This new switch challenges the notion that specialized protocols like Ultra Accelerator Link (UALink) or Nvidia's NVLink are necessary for scale-up computing in AI and HPC environments 1

1

2

Key Features and Performance

The Tomahawk Ultra boasts impressive specifications that put it in direct competition with specialized interconnects:

1. Ultra-low latency: Achieves 250 nanoseconds switch latency at full 51.2 Tbps throughput 2

   2

   .
2. High performance: Capable of processing up to 77 billion packets per second, even with minimum packet sizes of 64 bytes 1

   1

   .
3. Scalability: Supports scale-up systems with at least 1,024 accelerators, potentially doubling Nvidia's current capabilities 1

   1

   .
4. Lossless fabric: Implements Link Layer Retry (LLR) and Credit-Based Flow Control (CBFC) to eliminate packet loss and ensure reliability 2

   2

   .

Broadcom's Strategy: Scale-up Ethernet (SUE)

Broadcom is promoting a technology called scale-up Ethernet (SUE) as an alternative to specialized protocols. Pete Del Vecchio, product line manager for Broadcom's Tomahawk line, emphasized the benefits of using Ethernet for all parts of the network, including monitoring, telemetry, and debugging tools 1

1

Comparison with Competitors

The Tomahawk Ultra offers significant advantages over its competitors:

1. Bandwidth: At 51.2 Tbps, it provides nearly twice the bandwidth of Nvidia's fifth-gen NVLink switches (28 Tbps) 1

   1

   .
2. Latency: Broadcom claims better latency than UALink, although this is difficult to verify until UALink hardware ships 1

   1

   .
3. Compatibility: Pin-compatible with Tomahawk 5, allowing for easy upgrades to existing switch chassis 1

   1

   2

   2

   .

Related Stories

Broadcom's Tomahawk 6: A Game-Changer for AI Networking

Broadcom Unveils Tomahawk Ultra: A New Networking Chip to Accelerate AI Processing

Broadcom Challenges Nvidia with Thor Ultra: A New Era in AI Networking

Broadcom's Jericho4 Chip: Revolutionizing AI Infrastructure with Distributed Data Center Connectivity

Industry Implications

The introduction of Tomahawk Ultra could have far-reaching effects on the AI and HPC networking landscape:

1. Challenge to specialized protocols: Broadcom's approach questions the necessity of new protocols like UALink 1

   1

   .
2. Ethernet optimization: The switch features adaptable, optimized Ethernet headers, reducing overhead while maintaining compliance 2

   2

   .
3. Potential industry shift: Success of Tomahawk Ultra could lead to broader adoption of Ethernet in HPC and AI clusters.

Broadcom's Position in the Market

While Broadcom remains a member of the UALink consortium, it's clear that the company is betting on Ethernet's potential in the HPC and AI space. Robin Grindley, a principal product line manager at Broadcom, noted that other transport protocols like UALink or Infinity Fabric can be transported over Ethernet if necessary 1

1

As the AI and HPC markets continue to evolve rapidly, Broadcom's Tomahawk Ultra represents a significant development in networking technology. Its ability to compete with specialized protocols while leveraging the widespread adoption and tooling of Ethernet could potentially reshape the landscape of high-performance interconnects for AI and scientific computing.