Abstract

Artificial Intelligence (AI) places new and unprecedented demands on servers and the network fabric. During the AI training process, massive amounts of data need to be exchanged among servers. Hardware offload via the server NIC can provide highly efficient network data transport and reduce AI training run times while freeing server resources, such as CPU, for other tasks. 

Dell commissioned Tolly to benchmark the IPv6 RDMA offload performance of the Broadcom 57508 Dual-Port 100 Gb/s Ethernet Controller (LAN adapter) running in the Dell PowerEdge R760xa Rack Servers by testing with Keysight Technologies AI benchmarks. Testing focused on evaluating the network throughput of Broadcom hardware offload as an alternative to traditional, software-based, network server stack processing.  Systems were connected across a spine-leaf network of Dell PowerSwitch Z-Series Ethernet switches. 

Tests showed that RDMA hardware offload was able to achieve over 97% of the theoretical maximum network throughput across the servers tested.