Abstract

The Tolly Group evaluated the Dell PowerSwitch Z9864F-ON in a Dell Networking AI fabric designed to address one of the biggest challenges in Ethernet-based AI environments: congestion caused by bursty, latency-sensitive training traffic. The report focuses on Dell Enterprise SONiC Distribution and its proprietary RoCE prioritization, which dynamically identifies and protects AI traffic when the fabric becomes oversubscribed.  

Tolly built a deliberately oversubscribed test environment using three Dell PowerSwitch Z9864F-ON switches in a spine-leaf topology. Two leaf switches connected to a spine switch using three 800GbE uplinks each, providing 2.4Tbps of aggregate bandwidth. Background traffic generated by a Keysight Ixia AresONE-S-800GE system was configured to consume 100% of available bandwidth using IPv4 traffic with 4096-byte packets at line rate. Synthetic AI/NCCL-like traffic was then introduced as additional RoCE traffic to test whether the fabric could preserve AI performance under congestion.  

With Dell’s congestion mitigation enabled, the switch dynamically reduced one stream of background traffic from 100% to 96%, freeing bandwidth for the AI flow. The result was a successful AI “all-to-all” collective that completed in approximately 11 seconds, delivered 94.42GB/s of AI bandwidth, and recorded zero retransmitted frames. When the same test was repeated with congestion mitigation disabled, the AI session failed and was aborted due to timeouts and exceeded retries.  

The report also highlights support for the linear pluggable optics ecosystem, including DR8-LPO, DR4-LPO, and VR4-LPO links in switch-to-switch and server-to-switch deployments. Tolly notes that LPO technology can reduce latency, power use, and complexity by shifting DSP functions into the switch ASIC. Overall, the evaluation presents the Z9864F-ON as an 800GbE AI fabric switch optimized for both congestion-sensitive AI traffic and next-generation optical connectivity.