Distributing AI workloads across multiple servers with GPUs, while using DPUs (e.g., NVIDIA BlueField) to manage network and storage tasks, best improves efficiency and performance in an AI data center. This approach leverages GPU parallelism for computation and offloads networking/storage (e.g., RDMA, encryption) to DPUs, reducing CPU overhead and latency. NVIDIA's "BlueField DPU Documentation" and
"AI Infrastructure for Enterprise" highlight this as an optimized design for scalable, high-performance AI deployments.
Consolidating workloads on one server (B) creates a bottleneck and single point of failure. Assigning networking to CPUs (C) negates DPU benefits, reducing efficiency. NVIDIA's architecture guidance supports distributed GPU-DPU setups.