Abstract

H3C S12500CR is a next-generation flagship switch launched by H3C for intelligent computing  large-scale models and high-performance computing data center scenarios. Its hardware design adopts a CLOS+ orthogonal architecture, enabling rate convergence between network nodes and computing nodes, providing a 100% lossless data channel for networks and AI computing. It supports high-density, high-speed interface cards, meeting the requirements of ultra-large-scale data centers and AIGC computing networks for non-blocking high-density server access.

Tolly conducted tests evaluating the performance of the NVIDIA Collective Communication Library (N CCL) and the large-scale model Llama3 across different network architectures with 64 GPUs. Specifically, the test compared the performance differences between the RoCE network using the H3C S12508CR switch and the InfiniBand (IB) network using the NVIDIA QM9700 switch. The IB network followed the multi-track topology shown on the right side of Figure 1, while in the RoCE network, the H3C S12508CR switch directly connected all servers.

The test results for NCCL and the large language model Llama3 demonstrate that, under the same workload scenarios, RoCE delivers performance comparable to IB and provides a consistent user experience.