In a move aimed at breaking Nvidia’s dominance, Broadcom has unveiled a new networking processor, the Tomahawk Ultra, designed to accelerate AI data processing. This chip, which uses an enhanced version of Ethernet, can connect four times as many chips as Nvidia’s competing technology.
Introduction: A new battle in the AI arena
Broadcom’s chip unit on Tuesday unveiled a new networking processor aimed at accelerating artificial intelligence data processing, which requires linking hundreds of chips together to act as a single entity. This new chip is the latest piece in Broadcom’s arsenal targeting its giant AI competitor, Nvidia. This step comes at a time when Broadcom is already helping Alphabet’s Google produce its own custom AI chips, which developers and industry experts consider one of the few viable alternatives to Nvidia’s powerful graphics processors.
Tomahawk Ultra: The ultimate data traffic cop
Broadcom’s new chip, dubbed Tomahawk Ultra, acts as a traffic cop for data moving quickly between dozens or hundreds of chips located close to each other within a data center, such as inside a single server rack. This role is critically important in AI applications, where communications between processing units must be extremely fast to ensure there are no bottlenecks hindering performance.
The chip aims to compete with Nvidia’s NVLink Switch chip, which serves the same purpose. However, Ram Velaga, a senior vice president at Broadcom, explained that the Tomahawk Ultra outperforms its competitor by being able to connect four times as many chips. Crucially, it does not use a proprietary data transfer protocol like Nvidia does, but instead uses a speed-enhanced version of the open Ethernet standard.
Ethernet versus NVLink: The battle of standards
The use of Ethernet represents a key selling point for Broadcom. Nvidia’s NVLink protocol is proprietary technology, meaning it works best within Nvidia’s closed ecosystem. This can force customers to purchase more Nvidia products to ensure compatibility and optimal performance.
In contrast, Ethernet is an open standard widely used across the networking industry. By offering a high-speed version of Ethernet tailored specifically for AI, Broadcom hopes to attract customers who prefer the flexibility and interoperability that open standards provide. This gives data center builders greater freedom to choose and mix components from different vendors.
Advanced development and manufacturing
Taiwan Semiconductor Manufacturing Company (TSMC) will manufacture the Ultra processor line using its advanced 5-nanometer manufacturing process. The design took Broadcom’s engineering teams nearly three years to develop. The chip was originally built for a market segment known as high-performance computing, but with the generative AI boom, Broadcom adapted the chip for use by AI companies because it is well-suited for scaling requirements.
Conclusion: Opening the door to competition
Both companies’ chips help data center builders and others link as many chips together as possible, a technique the industry calls “scale-up” computing. By ensuring that closely situated chips communicate quickly, software developers can harness the computational power required for AI.
The launch of the Tomahawk Ultra is an important step in the battle to provide the necessary infrastructure for the AI era. By offering a powerful and open alternative to Nvidia’s technology, Broadcom is not only opening the door to greater competition, but also giving customers more choices and flexibility in building the next generation of super AI systems.
