Tolly Test Report: Dell PowerEdge and Broadcom AI Performance Benchmarking
报告
Network stack processing can consume significant amounts of server resources. This is especially relevant to server systems running AI training models as that task consists of non-stop movements of data from server-to-server via the network. Offloading network stack processing to the server’s Ethernet controller (also known as a LAN adapter or network interface card (NIC)) offers an opportunity to reclaim server resources for AI and other application tasks.
The goal of this test was to demonstrate that hardware offload of the network stack provided by the Broadcom 57508 Dual-Port 100 Gb/s Ethernet Controller provided an effective, high-performance alternative to the traditional, software-based network stack implemented on the server. More specifically, the test focused on determining the benefit of offload in AI training scenarios.
To prove the capabilities, test engineers turned to Keysight — using the Keysight AI (KAI) Data Center Builder. This solution was chosen by Tolly because it can emulate the complex data flows generated by AI training - also known as collectives. While the server systems used in the test had GPUs installed, Keysight can emulate the network data flows generated by actual AI training processing without requiring the use of, or even the presence of, a GPU.
Thus, the KAI Data Center Builder significantly simplifies the process of benchmarking AI network performance. Read this report for the full test results.