Due to business needs, we may rent servers from AutoDL for model training, so we tested the performance of multi-GPU interconnect for consumer-grade graphics cards (original 2080Ti 11G) on their platform. The results are placed here for reference.
TL;DR: No NVLink. Although running on a PCIe 3.0x16 switch and both CPU NUMA nodes are connected to the PCIe switch, P2P communication is unavailable due to driver issues, resulting in poor multi-GPU interconnect performance.
2025/6/5About 13 min