Resource‑Decoupled Data Center Architecture and Emerging Technologies (DPU, IPU, CXL)

Traditional data center architectures rely on servers as the basic deployment unit, tightly coupling CPU, memory, GPU, and storage via internal buses, which leads to limited scalability, low resource utilization, and insufficient elasticity for emerging workloads like serverless computing and distributed training.

Resource‑decoupled data center designs break these physical boundaries by constructing separate pools of heterogeneous resources (CPU, GPU, FPGA, RAM, SSD, HDD) that are interconnected through high‑performance networks, enabling flexible allocation and higher utilization.

The evolution toward resource decoupling is driven by three factors: (1) diversified application requirements that favor specialized compute chips (e.g., AI workloads needing matrix operations), (2) the need for high‑speed network connectivity to replace intra‑server communication, and (3) increasingly powerful hardware control mechanisms that allow pooling of remote resources.

Three technical routes are identified for building resource‑decoupled data centers: CPU‑centric, memory‑centric, and fully decentralized architectures.

1. CPU‑centric approach keeps most processing on CPUs while offloading specific functions to DPUs such as Fungible DPU, Intel IPU, and Alibaba CIPU. The Fungible DPU (F1) combines data clusters, control clusters, and a network unit, offering 800 Gbps bandwidth, support for TCP/UDP, RDMA, and programmable P4 pipelines. Intel IPU offloads infrastructure tasks from CPUs, improving efficiency for cloud service providers. Alibaba’s CIPU integrates compute, storage, and networking acceleration, supporting AI frameworks like TensorFlow and PyTorch.

2. Memory‑centric and decentralized approaches leverage technologies like CXL (Compute Express Link), which provides three protocols—CXL.io, CXL.cache, and CXL.mem—to enable low‑latency, high‑bandwidth communication between CPUs, GPUs, FPGAs, and memory expansion devices. CXL 3.0 adds support for leaf‑spine topologies and port‑based routing, allowing up to 4096 nodes to be pooled across multiple racks.

Overall, resource‑decoupled architectures, powered by DPUs, IPUs, and CXL, promise higher resource utilization, better scalability, and the ability to meet the diverse performance demands of modern cloud, AI, and high‑performance computing workloads.