How Baidu Builds a Unified Edge‑to‑Cloud Distributed Architecture

Background and Motivation

Edge computing moves compute, storage, and networking capabilities from centralized data centers to the physical locations where data is produced and consumed. Baidu Intelligent Cloud offers a full suite of edge services—including globally distributed edge nodes (BEC), local compute clusters (LCC), and on‑premise edge servers (ECS)—to meet diverse customer proximity requirements.

Observed Market Trends

Four key changes have emerged from ongoing collaborations with customers:

Increasing demand for seamless integration of edge and central cloud resources.

Growing depth of scenario penetration, shifting from pure IaaS to edge‑focused PaaS middleware.

Escalating resource fragmentation as customers manage dozens to hundreds of geographically dispersed compute sites.

Rising diversity of workload requirements, necessitating flexible compute and networking capabilities.

These trends drive the transition from isolated edge solutions to a unified distributed cloud.

Four Critical Architectural Paths

“One Cloud, Multiple Chips” to support differentiated workloads such as cloud‑gaming with sub‑20 ms end‑to‑end latency.

Same‑source, same‑stack design to ensure consistent experience and advanced capabilities across central and edge clouds.

Hyper‑converged, miniaturized infrastructure that balances scale, elasticity, and cost.

Full cloud‑native transformation to simplify management of wide‑area distributed resources.

Key Technical Implementations

One Cloud, Multiple Chips : Baidu deployed proprietary ARM‑based “Panyu Hive” servers at edge nodes, combined with lightweight containers, to dramatically increase deployment density for cloud‑gaming. PC‑FARM servers with hardware virtualization and end‑to‑end kernel tuning enable high‑performance Windows cloud‑gaming. Custom GPU kernel isolation allocates GPU resources on demand, while hybrid networking provides tenant isolation and seamless inter‑operation between ARM and x86 instances.

Same‑Source, Same‑Stack Distributed Cloud OS : Instead of installing a full virtualization control plane on every edge node (which would be costly), Baidu introduced a cloud‑edge channel coordinator and eventual‑consistency control semantics, achieving a decoupled control and data plane that maintains transparent, uniform management across the wide‑area fabric.

Middleware Transparency : The cloud‑edge channel also enables transparent deployment of middle‑ware services (e.g., distributed RDS and SCS) so customers can manage both central and edge middleware from a single console.

High‑Bandwidth Cloud‑Intelligent Network (CSN) : Leveraging Baidu’s 100 Tbps backbone, CSN provides large‑bandwidth, low‑latency links between edge and central nodes, with automatic route learning and multi‑link optimization.

Auto‑Driving Data Ingestion : By routing vehicle data through city‑level fiber PoPs on the backbone, Baidu reduces cost and latency compared with traditional hard‑disk shipment, and implements protocol‑level optimizations for near‑edge data upload and command delivery.

Hyper‑Converged, Small‑Scale Gateway : Using smart‑NIC hairpin offload, Baidu built a zero‑CPU‑cycle gateway capable of 200 GB throughput and 3000 Mpps, deployed across public edge gateways, DDoS scrubbing, and dedicated line gateways.

Cloud‑Native Serverless Edge Engine : Based on virtual node technology, Baidu offers a serverless container engine that abstracts away cloud‑edge coordination, allowing customers to run workloads on a single Kubernetes‑compatible cluster. The engine supports heterogeneous accelerators (GPU, TPU, VPU), distributed storage, and VPC‑level networking unified by CSN.

Conclusion

Baidu Intelligent Cloud’s distributed cloud architecture combines edge‑to‑cloud integration, same‑source stack consistency, hyper‑converged hardware, and cloud‑native automation to address the scalability, cost, and operational challenges of modern edge workloads, ranging from cloud gaming to autonomous‑driving data pipelines.