Broadcom has announced the introductions of a new on-chip neural network inference engine called NetGNT (Network Generalized Neural Network  Traffic Analyzer) in its new software-programmable Trident 5-X12 chip.

NetGNT works in parallel to enhance the standard packet processing pipeline. The standard pipeline is single-packet/single-path, meaning it looks at one packet as it traverses the ports and buffers of the chip. In contrast, NetGNT is a n ML inference engine that can be trained to identify different types of traffic patterns across the entire chip.

Trident 5-X12 can be software-programmed, field-upgraded, and provides a bandwidth of 16.0 Terabits/second, which is twice that of Trident 4-X9. It adds support fo 800G ports, allowing direct connections to Tomahawk 5. It supports 48x200G downlink ports and 8x800G uplink ports for 1RU data center Top of Rack (ToR).

• Compared to Trident 4-X9, the power consumption per 400G port has been reduced by 25%.
• It supports 800G port using Broadcom's 100G-PAM4 SerDes, enabling compatibility with up to 4m DAC and linear optical devices.
• Supports the transition from 400G NIC to next-generation CPU and GPU servers.
• Introduces NetGNT for enabling new features, enhancing network efficiency, and performance.
• Enhanced telemetry features allow deeper real-time insights into network operations, which can be used for training NetGNT.
• Increased forwarding/policy scale to support continuously growing network scales.
• Maintains compatibility with traditional APIs of current Broadcom switch chips.
• Fully programmable using Network Programming Language (NPL), allowing on-the-fly upgrades for adding new telemetry, security, and traffic engineering features.
• NPL is compatible with the current Trident 4 seriess, extensively used by customers for developing their custom switch applications.
• Supports enterprise-grade SONiC and SAI for seamless integration into data center operational frameworks.