Azure Stack Edge¶

Scope¶

Azure hardware-as-a-service edge appliances for compute, AI inference, and data transfer at the edge. Covers Pro 2, Pro GPU, and Mini R SKUs, Kubernetes/IoT Edge runtime, local and cloud-tiered shares, GPU-accelerated inference, and network function virtualization.

Purpose-built edge appliances delivered as a service (hardware-as-a-service) for running compute, AI inference, and data transfer workloads at the edge. The product line includes Azure Stack Edge Pro (with GPU or FPGA acceleration) for standard edge deployments and Azure Stack Edge Mini R (ruggedized, portable) for harsh or field environments. Azure manages the device lifecycle; customers order through the Azure portal, receive the hardware, activate it, and pay a monthly fee.

Checklist¶

Why This Matters¶

Edge locations — retail stores, factory floors, oil rigs, field hospitals, military forward operating bases — need local compute and AI inference without depending on reliable cloud connectivity. Azure Stack Edge provides this as a managed appliance: Microsoft owns the hardware, handles replacements, and pushes updates. This eliminates the operational burden of procuring, racking, and maintaining edge servers in potentially hundreds of locations.

The hardware-as-a-service model converts what would be a CapEx hardware purchase into a predictable monthly OpEx charge, which simplifies procurement in organizations where capital budgets are constrained. However, the monthly fees accumulate over multi-year deployments, so total cost of ownership should be compared against self-managed edge infrastructure for long-duration use cases.

AI inference at the edge is the primary differentiator. Rather than streaming raw video or sensor data to the cloud for processing (which consumes bandwidth and introduces latency), models run locally on GPU hardware with results sent to the cloud as structured data. This reduces bandwidth costs by orders of magnitude and enables real-time decision-making.

Common Decisions (ADR Triggers)¶

Azure Stack Edge Pro vs. Mini R¶

Pro devices are datacenter/closet-grade 1U appliances with higher compute and storage capacity, suitable for permanent installations. Mini R is ruggedized (IP65-rated), battery-capable, and portable, designed for field operations, disaster response, or deployments in harsh environments. Record the deployment environment and mobility requirements to justify the SKU selection.

GPU vs. FPGA SKU¶

FPGA SKUs were retired in February 2024. New deployments must use GPU SKUs. If migrating from FPGA, document the model conversion plan (FPGA models to ONNX/GPU-compatible formats) and timeline.

Kubernetes (Arc-enabled) vs. IoT Edge runtime¶

Both run containers on the device. Kubernetes (Arc-enabled) provides standard K8s APIs, Helm chart deployment, and GitOps via Flux. IoT Edge provides module-based deployment managed through IoT Hub, with built-in message routing and device twin capabilities. Choose Kubernetes for cloud-native teams and workloads that may migrate between edge and cloud. Choose IoT Edge for IoT-centric scenarios with device management, message brokering, and integration with IoT Hub analytics.

Edge local shares vs. cloud-tiered shares¶

Edge local shares keep data on the device only — suitable for scratch space, temporary processing data, or air-gapped scenarios. Cloud-tiered shares automatically upload data to Azure Storage — suitable for data collection, backup, and analytics pipelines. Document the data residency and data flow requirements for each share.

Self-signed vs. CA-issued certificates¶

Self-signed certificates work for initial testing but cause trust warnings and complicate automation. CA-issued certificates (from internal PKI or public CA) are required for production to enable trusted connections from client machines and automated systems. Document the certificate lifecycle and renewal process.

Single device vs. multi-device deployment at scale¶

For deployments spanning many sites (retail chains, cell towers), use Azure Resource Manager templates and Azure CLI/PowerShell for repeatable provisioning. Consider Azure Lighthouse for multi-tenant management if an MSP manages the devices. Record the scale-out strategy and per-site configuration management approach.

Bandwidth scheduling for data transfer¶

Cloud-tiered shares can saturate WAN links. Bandwidth schedules throttle or pause uploads during business hours. Document the bandwidth allocation, schedule windows, and priority of data transfer relative to other site traffic.

Reference Architectures¶

Retail edge — AI-powered inventory and checkout¶

Azure Stack Edge Pro 2 (single GPU) deployed in a store backroom. Computer vision models (ONNX format, deployed via Azure ML) process camera feeds to detect shelf stock levels and checkout anomalies. Edge local shares store video buffer temporarily; structured inference results (item counts, alerts) are sent to Azure IoT Hub for aggregation. Cloud-tiered shares upload daily summary data to Azure Blob Storage for retraining pipelines. IoT Edge runtime manages camera ingestion modules and inference modules. Azure Monitor tracks device health; store IT staff interact only with the Azure portal.

Manufacturing — defect detection on production lines¶

Azure Stack Edge Pro 2 (dual GPU) at the factory edge. High-resolution camera feeds from production lines are processed by defect detection models running on the GPUs. Inference latency must be sub-100ms to enable real-time rejection. Results are published to a local MQTT broker (running as an IoT Edge module) and consumed by the PLC/SCADA system. Cloud-tiered shares upload defect images and metadata to Azure Blob for model retraining in Azure ML. OPC UA adapter modules connect to industrial equipment for telemetry collection.

Field operations — disconnected/semi-connected¶

Azure Stack Edge Mini R carried in a transit case to field locations (disaster response, military, geological survey). Device operates on battery or generator power. Kubernetes workloads run data processing pipelines entirely offline. Data is stored on edge local shares. When connectivity is available (satellite link, temporary cellular), cloud-tiered shares sync collected data to Azure. Azure Arc-enabled Kubernetes allows remote workload updates when connected. Tamper-resistant enclosure and BitLocker encryption protect data in transit and at rest.

Telco edge — Azure Private 5G Core¶

Azure Stack Edge Pro deployed at cell tower sites or central offices. Azure Private 5G Core (packet core) runs as a network function on the device, providing private LTE/5G connectivity for enterprise campuses. Kubernetes hosts the packet core and additional edge applications (video analytics, AR/VR). The device connects back to Azure for management and monitoring. Multiple devices across sites are managed as a fleet through Azure portal and ARM templates.

Data transfer and migration — Data Box Gateway mode¶

Azure Stack Edge configured primarily as a data gateway. NFS/SMB shares are mounted by on-premises servers and applications. Data written to shares is automatically transferred to Azure Blob or Azure Files. Bandwidth schedules control transfer windows. This architecture replaces physical Data Box shipments for ongoing data ingestion from sites with adequate network connectivity. Useful for media production (dailies upload), scientific data collection, and backup-to-cloud workflows.

Reference Links¶

Azure Stack Edge documentation -- device setup, activation, compute configuration, and data transfer
Azure Stack Edge Pro 2 specifications -- hardware specs, GPU options, and supported workloads
Azure Stack Edge Mini R overview -- ruggedized edge appliance specifications and deployment scenarios