Edge AI Data Center with RDMA over All-Photonic Network

The architecture of the edge AI data center with remote GPU cluster is proposed to support the growing edge AI applications. This architecture leverages RDMA, the AllPhotonic Network (APN), and the remote GPUs to provide powerful computing resources in the edge AI data center. In addition, we also introduce a RDMA accelerator with a cache preloading mechanism to enhance data transfer efficiency and reduce synchronization bottlenecks. The pilot’s results show that this architecture and mechanism improved data leakage and latency issues. For long-distance data transfers, our proposed architecture-NFS over RoCEv2 over APN with a cache preloading mechanism-significantly outperforms the traditional NFS over TCP without APN. When transferring a single large file, the traditional approach takes 62.28 seconds, while our architecture completes the task in just 9.83 seconds, achieving an $\mathbf{8 4. 2 2 \%}$ reduction in transfer time. Similarly, for a batch of $\mathbf{5 0, 0 0 0}$ small files, the traditional method takes $\mathbf{3 1 0. 5 2}$ seconds, whereas our solution reduces the time to $\mathbf{1 2 1. 3 6}$ seconds, yielding a $61 \%$ improvement in performance.