Why Hyper-Converged Data Center Networks Are the Future of AI-Driven Infrastructure

Introduction

Data‑center networks connect compute, storage, and high‑performance computing resources, and all server‑to‑server data exchanges must pass through the network. Rapid changes in IT architecture, compute, and storage are pushing data‑center networking from multiple isolated networks toward a fully Ethernet‑based, hyper‑converged model.

Why Hyper‑Converged Data Center Networks?

Traditional Ethernet cannot meet the stringent requirements of storage and HPC workloads. A hyper‑converged network uses lossless Ethernet to host general compute, storage, and HPC on a single fabric, enabling full‑life‑cycle automation and intelligent operations.

Current Situation: Three Networks Inside a Data Center

InfiniBand for HPC workloads

Fiber Channel for storage

Ethernet for general compute

AI‑Era Changes

AI workloads generate massive data, making network latency a critical bottleneck. Storage has shifted from HDD to SSD, reducing latency by over 100×, while compute has moved to GPUs and AI chips, boosting processing speed by more than 100×. Consequently, network latency now accounts for 60%+ of end‑to‑end delay.

RDMA Replaces TCP/IP but Has Limitations

TCP/IP adds tens of microseconds of fixed latency, which is unacceptable for AI and SSD‑distributed storage. RDMA reduces data‑transfer latency to ~1 µs and offloads CPU, but existing RDMA deployment options—InfiniBand and traditional IP Ethernet—have drawbacks such as closed architectures, high OPEX, or intolerable packet‑loss sensitivity.

Distributed Architecture Increases Congestion

Widespread adoption of distributed systems creates massive inter‑server traffic, large packets, and incast flows that trigger congestion and packet loss, further stressing the network.

Core Metrics for Hyper‑Converged Networks

The next‑generation network must achieve three inter‑dependent goals: zero packet loss, ultra‑low latency, and high throughput. Meeting all three simultaneously requires sophisticated congestion‑control algorithms.

Difference from HCI

Hyper‑Converged Infrastructure (HCI) integrates compute, storage, and networking in a single appliance, requiring extensive re‑architecting of resources. In contrast, a hyper‑converged data‑center network focuses solely on the network layer, leveraging Ethernet for low‑cost, rapid scaling without altering compute or storage stacks.

Huawei’s Hyper‑Converged Network Solution

Huawei combines years of data‑center networking experience with an iLossless intelligent lossless algorithm that dynamically adjusts network parameters for zero‑loss operation. The solution uses a CLOS spine‑leaf architecture built on CloudEngine switches, integrating compute‑intelligence and network‑intelligence for global and local optimization.

Huawei’s iMaster NCE‑FabricInsight platform collects traffic features and network state, applies AI models to predict future traffic, and automatically tunes NIC and network settings.

Value Proposition

Performance Boost : Up to 44.3% reduction in compute latency for HPC and 25% IOPS improvement for distributed storage, while guaranteeing zero packet loss.

Cost Reduction : Network accounts for only ~10% of data‑center CAPEX; the solution leverages existing Ethernet skills, delivering up to 10× leverage on server/storage investment and significant ROI.

SDN Automation & Intelligent Operations : Supports full‑life‑cycle SDN automation, cutting OPEX by over 60% and enabling visual, multi‑dimensional network management via AI‑driven analytics.

How It Works

Using RoCEv2 over lossless Ethernet, Huawei implements three complementary technologies:

Traffic‑Control : PFC deadlock detection and prevention to avoid congestion‑induced packet loss.

Congestion‑Control : AI‑enhanced ECN, intelligent QCN, ECN Overlay, and Network‑Based Proactive Congestion Control (NPCC) to maintain high throughput.

Intelligent Lossless Storage : iNOF (Intelligent Lossless NVMe‑over‑Fabric) provides rapid host‑side control for storage traffic.