FILE PHOTO: Broadcom Limited company logo is pictured on an office building in Rancho Bernardo, California May 12, 2016. REUTERS/Mike Blake/File Photo
(Reuters) - Tech conglomerate Broadcom added artificial intelligence features to a new version of one of its flagship networking chips, the company said on Thursday, aiming to help move information around data centers more efficiently.
Devoting a portion of the silicon in Broadcom's Trident networking processors to AI improves the chip's efficiency and its ability to handle other tasks such as network security. The new chip includes typical performance boosts such as reduced power consumption and more network bandwidth.
One of the most significant difficulties when building massive AI clusters is moving data around inside them.
The chip, called a Trident 5-X12, can help alleviate some of the network traffic congestion, Robin Grindley, an executive in Broadcom's Core Switching Group, said in an interview. There are specific networked computing tasks, such as AI, that can only be improved by adding additional capabilities to a chip, because software is too slow.
"That's what the neural network does - it looks across all packets, all traffic patterns, so it's trying to identify these things that the standard approach just wouldn't be able to catch," Grindley said.
The AI portion of the chip can be turned on after a customer builds an AI model based on the traffic in their data center. Data center operators can train a model to sniff out specific types of unwanted traffic such as denial of service attacks or network traffic congestion. Once constructed, a form of the model is fed into the chip and it can help route the traffic.
Engineers at the company decided to add the AI features roughly two years ago after the company introduced programmability into a prior version of the chip, Grindley said.
The new version of the Trident is built with 5 nanometer manufacturing technology and is currently shipping to "qualified" customers.
(Reporting by Max A. Cherney in San Francisco; Editing by Sonali Paul)
