Broadcom has introduced a new networking chip designed to boost AI data centre performance, positioning itself against Nvidia in the AI hardware race.
Called Tomahawk Ultra, the chip helps link hundreds of processors within close range, ensuring fast communication vital for large AI models.
Instead of using Nvidia's proprietary NVLink system, Broadcom's Tomahawk Ultra operates on an accelerated version of Ethernet and connects up to four times more chips within a server rack.
The chip acts as a traffic controller, enabling what the industry refers to as 'scale-up' computing - where AI models tap into combined computing power from tightly grouped chips.
According to Broadcom senior vice president Ram Velaga, it took engineers around three years to design the Tomahawk Ultra, which was originally aimed at high-performance computing but adapted for AI workloads as demand surged.
Taiwan Semiconductor Manufacturing Company (TSMC) is producing the processors using its five-nanometre process, and the chips are already shipping to customers.
Rather than requiring major hardware overhauls, the Tomahawk Ultra offers a scalable, Ethernet-based alternative for AI firms looking to build faster, more flexible data centres.
