Broadcom unveils new Tomahawk 6 network chip to accelerate artificial intelligence
Broadcom (AVGO.O) announced on Tuesday that it has begun shipping its latest network chip designed to accelerate AI-related computing.
The new chip, named Tomahawk 6, delivers twice the performance of its predecessor and includes improved traffic management features that significantly boost efficiency, according to Broadcom’s Senior Vice President Ram Velaga in an interview with Reuters.
Velaga said the increased speed reduces the number of network switches required to perform the same tasks.
Broadcom’s network chips are becoming increasingly important amid surging demand for AI infrastructure. To support these systems, infrastructure developers need to interconnect hundreds or even thousands of chips into a single network.
Building large data processing clusters requires specialized networking equipment, and this is where Tomahawk chips play a critical role. According to Velaga, Tomahawk 6 enables the creation of even larger AI data centers — including those connecting over 100,000 GPUs.
“In a few years, you’ll start to see a million GPUs located within a single physical building,” he said.
Broadcom’s chips use the Ethernet protocol, which has remained an industry standard for decades. Meanwhile, Nvidia (NVDA.O) promotes its own hardware based on the competing InfiniBand technology, along with some Ethernet-based solutions.
“All of these networks can be very easily built on Ethernet — there’s no need for esoteric technology,” Velaga added.
Tomahawk 6 is the first in its series to use multiple chiplets — individual chips combined in a single package — a design approach already widely adopted by other companies, including AMD. According to Velaga, the use of chiplets has doubled the amount of silicon used in the chip’s architecture.