By Max A. Cherney
SAN FRANCISCO (Reuters) -Broadcom has begun to ship its latest networking chip that aims to speed AI, the company said on Tuesday.
The chip, called the Tomahawk 6, boasts double the performance compared with the prior version and other traffic control features that make the networking chip significantly more efficient, Ram Velaga, a Broadcom senior vice president, told Reuters in a Monday interview.
The speed boost means that fewer networking switches are needed to perform the same task, Velaga said.
Broadcom’s networking chips have gained increased importance because of AI. When constructing the necessary data centers for AI applications, infrastructure builders must string together hundreds or thousands of chips.
Building large-scale clusters of networked chips requires specialized networking gear and chips, of which the Tomahawk series of processors is one such component.
With the Tomahawk 6, Broadcom’s engineers have boosted its speed and capabilities to the point where it can be used to construct the larger data centers that are necessary for AI, which can be over 100,000 graphics processors (GPUs) strung together, Velaga said.
“In a couple of years, you will start to see a million GPUs housed inside a physical building,” he said.
Broadcom’s networking chips use the Ethernet networking protocol, which has been a networking standard for decades. Nvidia produces hardware that uses a rival tech called InfiniBand and several products based on Ethernet.
“All of these networks can be very simply done on Ethernet, you don’t need esoteric technologies,” Velaga said.
The Tomahawk 6 is the first product in that line that will use several chips combined into a single package, a tech known as chiplets that is widely adopted by other chip designers such as Advanced Micro Devices. Adding chiplets roughly doubled the amount of silicon area used in the design, Velaga said.
Broadcom is producing the Tomahawk switch on Taiwan Semiconductor Manufacturing Co’s three nanometer process.
(Reporting by Max A. Cherney; Editing by Leslie Adler)
Comments