Broadcom Inc. has introduced what it claims is the first switch with an on-chip, neural network inference engine.
Called Trident 5-X12, the software-programmable switch uses the networking general-purpose neural-network traffic analyzer (NetGNT) to double bandwidth and reduce power by 25% compared to previous devices.
The NetGNT technology works in parallel to augment the standard packet-processing pipeline. The technology is a machine learning inference engine and can be trained to look for different types of traffic patterns that span the entire chip, Broadcom said.
NetGNT can look at traffic patterns common in artificial intelligence and machine learning workloads called “incast” that calls many packet flows to converge on the same port and buffer at the same time, causing congestion. The technology recognizes this pattern in real-time and begins and invoke congestions-control techniques to avoid degraded network performance.
The Trident 5-X12 device:
- Adds support for 800G ports
- Supports the transition to next-generation CPU and GPU servers
- Adds enhanced telemetry capabilities
- Increases forwarding/policy scales
- Maintains legacy API compatibility
- Features full programmability using NPL
- Demonstrates compatibility with current Trident 4 families of devices
- Features Enterprise Sonic and SAI support