Broadcom's on-chip neural-network inference engine to improve n/w performance of switch
Date: 12/12/2023
Broadcom has integrated a new on-chip neural-network inference engine called NetGNT in its latest software-programmable Trident 5-X12 chip for network switch. This engine handles AI/ML load triggered traffic congestion such as “incast” where multiple packet flow and converge on the same port and buffer at roughly the same time causing congestion. NetGNT recognize such patterns in real-time and enable congestion-control techniques and up the network performance.
This ML inference engine runs in hardware and can be trained dynamically to look for different types of traffic patterns that span the entire chip without affecting throughput and latency. NetGNT works in parallel augmenting the standard packet-processing pipeline.
“We continue to push the envelope and introduce brand-new technologies such as NetGNT to the market. We also listen very closely to our customers who have made it clear that one size does not fit all. They rely on us to deliver a broad portfolio of chips, customized for different applications,” said Ram Velaga, senior vice president and general manager, Core Switching Group, Broadcom. “Trident 5-X12 is the most power-efficient ToR (Top of Rack) on the market, while still adding cutting-edge new features that our customers have come to expect from an innovator such as Broadcom.”
The software-programmable and field-upgradable Trident 5-X12 uses 25% less power per 400G port than Broadcom's popular Trident 4-X9. Trident 5-X12 supports 800G ports to build a 1RU data center ToR supporting 48x200G downlink ports and 8x800G uplink ports.
Other features listed by Broadcom includes:
Supports the transition to next-generation CPU & GPU servers using 400G NICs
Adds NetGNT to enable new capabilities along with improved network efficiency and performance
Adds enhanced telemetry capabilities allowing deeper real-time insights into network operations, which can then be used to train NetGNT
Increases forwarding/policy scales to support ever-expanding network sizes
Maintains legacy API compatibility with current Broadcom switch chips
Fully-programmable using NPL (Network Programming Language), allowing in-field upgrades to add features such as new telemetry, security and traffic engineering
NPL-compatible with current Trident 4 family, which is already in use by a large number of customers to develop their own customized switch applications
Support for Enterprise SONiC and SAI promote rapid and seamless integration into data center operations frameworks.
Broadcom is now shipping Trident 5-X12 devices to qualified customers
Source: Broadcom
