Broadcom’s Neural Network Inference Engine Sets Sail for Shipping

Broadcom has announced the release of a groundbreaking on-chip neural network inference engine called NetGNT (Networking General-purpose Neural-network Traffic-analyzer). This innovative technology is integrated into the new Trident 5-X12 chip, which is software-programmable and offers exceptional performance.

NetGNT operates in parallel with the standard packet-processing pipeline, enhancing its capabilities. Unlike the traditional one-packet/one-path approach, NetGNT is a machine learning (ML) inference engine that can be trained to identify various traffic patterns across the entire chip.

One of the key advantages of NetGNT is its ability to detect traffic patterns such as "incast," commonly found in AI/ML workloads. Incast refers to the convergence of multiple packet flows on the same port and buffer simultaneously, leading to congestion. NetGNT can identify this pattern in real-time and implement congestion-control techniques to prevent network performance degradation. Furthermore, as NetGNT operates in hardware at full line rate, there is no impact on throughput or latency.

The Trident 5-X12 chip, which incorporates NetGNT, offers impressive features and capabilities. It provides a bandwidth of 16.0 Terabits/second, double that of the market-leading Trident 4-X9. Additionally, it supports 800G ports, enabling direct connectivity to Broadcom's advanced Tomahawk 5, which serves as the spine/fabric in cutting-edge compute and AI/ML data centers.

This chip is specifically designed to facilitate the deployment of a 1RU data center Top-of-Rack (ToR) solution, supporting 48x200G downlink ports and 8x800G uplink ports. With its software-programmable and field-upgradable nature, the Trident 5-X12 chip offers flexibility and scalability for evolving network requirements.

For more information, click here.