Abstract

Dell Technologies commissioned Tolly to evaluate the Dell AI Factory solution with NVIDIA Spectrum with Dell SONiC. The solution includes Dell PowerSwitch SN5610 NVIDIA Spectrum with Dell SONiC for AI Fabric, and Dell PowerSwitch SN2201 NVIDIA Spectrum with Dell SONiC switches with Dell SONiC for management, in addition to NVIDIA GPUs and NVIDIA BlueField 3 SuperNICs and DPUs in the Dell PowerEdge XE9680 servers.

Dell’s AI Factory solution with NVIDIA Spectrum and Dell SONiC is positioned as an Ethernet-based AI fabric designed to deliver near line-rate performance for large-scale GPU clusters while preserving the operational flexibility of an open networking stack. In Tolly’s evaluation, the solution combined Dell PowerSwitch SN5610 NVIDIA Spectrum with Dell SONiC for the AI fabric, Dell SN2201 NVIDIA Spectrum with Dell SONiC for management, and Dell PowerEdge XE9680 servers equipped with NVIDIA H200 SXM GPUs plus NVIDIA BlueField-3 SuperNICs and DPUs.  

The report’s central finding is that the fabric achieved near theoretical maximum performance on 400GbE links in both synthetic and AI-relevant communication tests. In the RDMA microbenchmark, using RoCE write-bandwidth testing, the fabric delivered 392.18Gbps, equal to 98% of 400GbE line rate. In the NCCL AllReduce test using 32 GPUs, the environment achieved 389.84Gbps of NCCL bus bandwidth, representing 97.5% efficiency. Tolly characterizes these results as evidence that the network introduced no measurable performance overhead while supporting enterprise features needed for production AI infrastructure.  

The evaluation also emphasizes that these are not purely switch-level figures, but end-to-end measurements across the AI fabric under realistic distributed GPU communication conditions. The NCCL test is particularly significant because AllReduce is one of the most common collective operations used during AI training, making the reported bandwidth a practical indicator of how effectively the fabric supports synchronized multi-node training workloads. According to the report, the results indicate strong fabric utilization, predictable application-level throughput, and efficient scaling of GPU-to-GPU communication across nodes.  

The tested environment used Ubuntu 24.04.3 LTS, NVIDIA AI Enterprise 7.1 components including CUDA 13.0 and NCCL 2.29.7, and Dell SONiC SN4.5.1.1. Tolly also notes that testing was conducted on the Dell SN5600, but that the SN5610 is its functional replacement, using the same NVIDIA Spectrum-4 ASIC and identical front-panel port configuration, making the performance results directly applicable to both models. Overall, the report presents Dell SONiC on NVIDIA Spectrum as a viable, high-performance Ethernet fabric for scalable AI deployments.