Resource Decoupling Architectures in Modern Data Centers: CPU‑Centric, Memory‑Centric, and Decentralized Approaches

With rapid advances in networking and hardware, resource‑decoupling architectures have become a primary direction for future data centers due to high resource utilization and excellent hardware scalability. Three main technical routes are identified: CPU‑centric, memory‑centric, and resource‑decentralized.

CPU‑Centric approaches keep all compute, storage, and network logic on the CPU, accessing remote resources via network I/O, which can introduce bandwidth and latency overheads in large‑scale deployments.

Memory‑Centric designs move traditional OS memory management to the memory side, enabling independent memory expansion and shared access by heterogeneous compute units. Notable implementations include HP The Machine (featuring SoC, unified storage, memory pooling, and optical communication), the Scorpion project (ODCC 2.0 specifications for modular rack servers), and NetDAM, which attaches memory directly to Ethernet controllers to provide in‑memory and in‑network computing.

NetDAM achieves deterministic latency by eliminating PCIe DMA and cache‑coherency checks, offers optional reliable transmission, supports out‑of‑order execution with packet‑level addressing, and enables multipath transport via UDP segment routing, thereby reducing latency and maximizing bandwidth.

Resource Decentralization removes the central control of any single resource type, distributing management functions to the respective resource controllers and using message‑passing for coordination. This paradigm aims to achieve high elasticity, on‑demand scalability, and independent evolution of compute, storage, and network resources.

Overall, while each route improves certain aspects of data‑center design, they still exhibit limitations in scalability and flexibility, prompting the emergence of resource‑decentralized architectures as a promising direction for future data‑center evolution.