Broadcom Takes on Nvidia with the Newly Launched Tomahawk Ultra Ethernet Switch

Global technology firm Broadcom has launched the Tomahawk Ultra Ethernet switch and started shipping it, Yahoo Finance reported. The new switch is designed to improve AI workloads and support high-performance computing (HPC). 

High Capabilities

Broadcom says Tomahawk Ultra is designed to optimize Ethernet headers by lowering overhead while maintaining compliance. The AI ethernet switch can reduce overhead from 46 bytes to 10 bytes. This helps to increase network efficiency and facilitates specific customizations. 

“Tomahawk Ultra is a testament to innovation, involving a multi-year effort by hundreds of engineers who reimagined every aspect of the Ethernet switch. This highlights Broadcom’s commitment to invest in advancing Ethernet for high-performance networking and AI scale-up,” Broadcom Senior VP and General Manager Ram Velaga said. 

The Tomahawk Ultra Ethernet switch can achieve a 250 nanoseconds latency at 51.2 terabits per second of throughput. It’s capable of supporting line-rate switching of packets as small as 64 bytes while handling a maximum of 77 billion packets per second. 

Broadcom’s new switch is designed to work in AI clusters and HPC environments. The switch provides adaptable Ethernet headers and super-low latency that improve efficiency in AI model training and large-scale simulations. 

“AI and HPC workloads are converging into tightly coupled accelerator clusters that demand supercomputer-class latency — critical for inference, reliability, and in-network intelligence from the fabric itself,” Lead Semiconductor Analyst at Bloomberg Intelligence Kunjan Sobhani said.

Ideal for Data Centers

When coupled with scale-up Ethernet, Tomahawk Ultra supports sub-400-nanosecond communication latency in processing units. This includes transit time through the switch. Broadcom incorporated fabric technology in the new switch to prevent packet drops during data transfers. 

The company also used Credit-Based Flow Control and Link Layer Retry to enhance reliability. These components enhance link error detection and automatic packet transmission. These processes create a lossless Ethernet fabric suitable for AI infrastructure in data centers. 

“Demonstrating that open-standards Ethernet can now deliver sub-microsecond switching, lossless transport, and on-chip collectives marks a pivotal step toward meeting those demands of an AI scale-up stack — projected to be double digit billions in a few years,” Sobhani added.

Broadcom vs Nvidia Networking

The new Broadcom networking chip takes on AI giant Nvidia. The company helps Google to produce AI chips which industry experts and developers consider to be one of the viable alternatives to Nvidia’s AI chips. Broadcom’s chip serves as a traffic controller for data moving between hundreds of chips in  data centers. 

It seeks to compete with Nvidia’s NVLink Switch, which works similarly. However, the Tomahawk Ultra switch is capable of tying together up to four times the number of chips. It leverages Ethernet, which has been boosted for speed as opposed to the proprietary protocol to shift data. 

Both switches are designed to help data center developers to tie many chips together in close proximity- a technique that’s known as scale-up in computing. By supporting fast communication between chips, software developers can maximize the computing power  required in AI. 

Broadcom engineers spent around three years designing the Tomahawk Ultra chip. Originally, the switch was developed for the high-performance computing market segment. However, the company adapted the switch for use by AI firms due to its scaling-up capabilities as the generative AI industry expanded. 

Earlier this week, Broadcom canceled a $1 billion microchip plant investment deal in Spain after talks with the government collapsed. The cancellation triggered a 1% stock price drop.