Broadcom’s Tomahawk Ultra chip connects 4x more AI processors than Nvidia

Broadcom has launched the Tomahawk Ultra networking chip, designed to accelerate AI data processing by connecting hundreds of chips within data centers. The processor directly challenges Nvidia’s dominance in AI infrastructure, offering four times the connectivity capacity of Nvidia’s competing NVLink Switch chip while using enhanced ethernet protocols instead of proprietary systems.

Why this matters: The chip represents Broadcom’s strategic push to compete with Nvidia in the lucrative AI hardware market, particularly as companies like Google seek alternatives to Nvidia’s graphics processors for AI workloads.

Key technical advantages: The Tomahawk Ultra serves as a traffic controller for data moving between dozens or hundreds of chips positioned closely together in server racks.
• The chip can tie together four times more processors than Nvidia’s NVLink Switch, according to Ram Velaga, a Broadcom senior vice president.
• Instead of using proprietary protocols, it employs a speed-enhanced version of ethernet for data transmission.
• Taiwan Semiconductor Manufacturing is producing the processors using its five-nanometer process technology.

In plain English: Think of the Tomahawk Ultra as a sophisticated traffic light system for a busy intersection. Just as traffic lights coordinate the flow of cars from multiple directions, this chip coordinates the flow of data between hundreds of AI processors working together in a data center. The closer these processors are to each other (within a few feet), the faster they can share information and solve complex AI problems.

Development timeline: Broadcom’s engineering teams spent roughly three years developing the design, originally targeting high-performance computing applications.
• The company adapted the chip for AI applications as generative AI demand surged, leveraging its natural suitability for “scale-up” computing.
• The processor is now shipping to customers.

Strategic positioning: Both Broadcom and Nvidia are targeting the critical “scale-up” computing segment, where close proximity between chips enables rapid communication necessary for AI processing.
• Broadcom supports Alphabet’s Google in producing AI chips that industry experts view as viable alternatives to Nvidia’s powerful graphics processors.
• The competition centers on helping data center builders maximize the number of chips they can connect within a few feet of each other.

What they’re saying: “The chip aims to compete with Nvidia’s NVLink Switch chip which has a similar purpose, but the Tomahawk Ultra can tie together four times the number of chips,” Velaga told Reuters.