Could the software company take the lead from AWS in remote AI cloud services at the network edge? An update on a strategy with a long history.
As early as 2016, Microsoft took the market by surprise by launching Azure Stack. Objective of the American software giant: to give the possibility to its customers to install Azure at home. Pre-integrated with racks of 4 to 16 servers signed HPE or Dell EMC, the product includes the main services of the public cloud, from the provisioning of virtual machines and network or storage resources to serverless, and allows them to be consumed on demand via a subscription model. In 2019, Microsoft will enrich this first offer with a hyperconverged infrastructure under Windows or Linux. The Holy Grail of the private cloud. Called Azure Stack HCI, this solution, previously known as Windows Server Software Defined (and limited to Windows) thus combines the virtualization of computing, storage and networking. Announced in parallel, Azure Stack Edge completes the picture. This is undoubtedly the most innovative solution in the Stack range, and is built around a server optimised to run machine learning (ML) models locally, which have been trained on Azure. The challenge is obviously to move towards real-time processing by avoiding round trips to the cloud.
Under the hood, the server embeds an FPGA (Field Programmable Gate Array) processor of the Intel Arria 10 type, designed for model inference. Supported scenarios include image recognition and TensorFlow neural networks. On the front end, models can be created and trained through Azure Machine Learning on Microsoft’s public cloud, before being deployed in a software container. To accommodate data from local IT systems or connected objects, Stack Edge is equipped with 12 terabytes of storage. “The complete learning data sets can be transferred to it if needed to reform and improve the ML,” adds Microsoft. The appliance is also equipped with a gateway for private cellular networks with the ability to ingest multiple IoT streams simultaneously. As a result, Stack Edge is able to operate fully in disconnected mode.
“Azure Stack Edge provides location-based networks with proximity computing units that install on site.”
“Azure Stack Edge provides location networks with proximity computing units that install on site,” comments Xavier Perret, Azure director at Microsoft France. “Its application areas are numerous. For example, it can be used to detect anomalies or perform predictive maintenance on an industrial site via image recognition(from cameras, editor’s note) or to anticipate imponderables in stock management or the supply chain in retail. Overall, Azure Stack Edge targets production processes that need to be driven in real time without depending on network latency.”
Another use case cited by Microsoft is the example of a customer support center that requires the implementation of an on-site voice recognition system to route incoming calls.
Unlike Stack Hub and Stack HCI, which can be purchased off the shelf before consuming the services, Azure Stack Edge is full cloud. Presented as a full Azure service, the server is fully managed by Microsoft and is priced at €586.9 per month. Customers can order it directly from the Azure portal. Once installed and connected locally, Stack Edge is controlled via the Microsoft cloud console. Retaining full sovereignty over their data, customers benefit from all the Azure tooling to manage deployments, access control and orchestrate workloads. A bridge is also available to Redmond’s cloud storage services APIs, including Azure Blob Storage for block storage and Azure Files for file storage. “Depending on local processing needs and their increase in power, it is possible to connect several Stack Edge servers, and add more as they go along to create a mini data center,” says Neal Analytics, an American ESN expert in AI and data analytics, a partner of the publisher.
A cheaper offer
For the future, Microsoft is working on several developments. “Azure Stack Edge will soon support new computing capabilities, including virtual machines and Kubernetes clusters,” says Xavier Perret. “We also plan to enrich this offer with GPGPU processors of the Tesla Nvidia type designed for high performance computing. Last but not least, Satya Nadella’s group is currently designing a reinforced version of Azure Stack Edge for remote missions, on construction sites or at risk industrial sites in particular. It will be battery-powered and will be able to nestle into a bag. “Following natural disasters for example, its machine learning capabilities can help find a missing person(via computer vision algorithms, ed.),” said Dean Paron, partner director of product management for Azure Storage and Edge (see video). No specific date has yet been given for this roadmap.
Facing Microsoft, Amazon Web Services (AWS) is currently the only cloud to market an equivalent of Azure Stack Edge. Called Outpost, it was launched at its latest annual customer event, re:Invent, held in December 2019. Like Stack Edge, this solution comes in the form of a 100% managed server infrastructure.
AWS Outpost runs Amazon’s Kubernetes orchestrator (EKS) on top of its container as a service (ECS) environment. But also Amazon RDS on the database side and Amazon EMR for distributed processing. The latter also offers the possibility of choosing Apache Spark and its machine learning library MLlib.
In addition to Amazon EBS storage capacities, Outpost’s portfolio includes a wide variety of EC2 instances, including GPU machines (G4dn) designed to run ML models. All of these computing resources are, of course, subject to additional pricing based on the AWS pricing model. This pay-as-you-go pricing is in addition to the initial cost of Outpost, which is marketed as amonthly subscription (entry price: $8,000 per month) or via a one-time payment at the time of ordering (entry price: $270,644). There’s no contest: Azure Stack Edge is much cheaper.