Marvell Technology

As AI clusters scale to tens of thousands of GPUs, Ethernet is no longer just a connectivity layer. It becomes a tightly coupled component of the compute system, where a single dropped packet can stall multiple GPUs and degrade overall utilization.

Marvell and Keysight Technologies demonstrated how the Marvell Teralynx switch fabric supports emerging Ultra Ethernet Consortium capabilities designed specifically for AI workloads: packet trimming, which preserves forward progress during congestion rather than triggering widespread retransmissions; Auto Load Balancing, which dynamically steers traffic away from congested links at line rate; and Ultra Ethernet Transport, which provides the framework within which both operate.

Senior Engineer Vikram Dattatri details the technology and what rigorous system-level validation under realistic AI workload conditions demonstrates about its performance at scale.

Learn more: https://mrvl.co/4yr2mMW