Broadcom is taking the wraps off new silicon today that will integrate an on-chip AI neural network designed to help improve networking.

The new Trident 5-X12 switching silicon features an integrated neural network engine called NetGNT (Networking General-purpose Neural-network Traffic-analyzer). According to Broadcom, this marks the first switch chip to contain an on-chip neural network, opening the door to more intelligent and efficient networks.

The new chip is an evolution of Broadcom's Trident silicon, which it has been iterating on for network connectivity and security for over a decade. The Trident 5-X12 provides 16.0 Tbps of bandwidth, double that of the previous generation Trident 4-X9.  The new Trident is also more power-efficient than its predecessor consuming 25% less power per 400G port.

The chip supports 800G ports for connecting directly to Broadcom’s Tomahawk 5 spine switches. This enables 1RU (rack unit) Top of Rack (ToR) switches with up to 48x200G downstream ports and 8x800G upstream ports.

Robin Grindley, principal PLM in Broadcom's  Core Switching Group, told SDxCentral that the new chip is the first member of the new Trident 5 family. The target deployment for the new silicon is in top of rack networking gear for both cloud and enterprise deployments.

At the heart of the Trident 5-X12 is the NetGNT neural network engine. NetGNT runs in parallel to the standard packet processing pipeline, analyzing traffic patterns in real-time.

For instance, NetGNT is built to detect impending network congestion and invoke congestion control mechanisms before performance degrades. This hardware-accelerated intelligence improves efficiency without impacting throughput or latency.

The NetGNT powered traffic analysis is different from the traditional types of analysis that Broadcom has previously enabled.

"At a high level, neural networks are good at looking for patterns across time and packets," Grindley said. "Traditional packet-processing looks for fixed criteria on a per-packet basis."

The NetGNT technology is not static and can be trained and updated. Grindely said that network packet traces need to be gathered, typically by the NOC (network operations center) in an organization. The packet traces are used for training, which is supervised. He added that NetGNT is loaded once inferencing weights are available and new weights can be loaded at any time.

Broadcom designed the Trident 5-X12 to be fully software-programmable using its NPL language. This is designed to allow customers to customize the chip with new features, protocols and capabilities through simple software upgrades.

The chip also comes with expanded telemetry capabilities. Deep traffic statistics and insights generated by the chip can help fine-tune NetGNT and algorithms to optimize network performance.While the Trident 5 and NetGNT are new technologies, Broadcom said that it  has positioned the Trident 5-X12 to be easily integrated with  existing switch software from Broadcom as well as supporting the open source based Enterprise SONiC network operating system.