Mirantis Adds AI Reference Architecture Based on k0rdent Control Plane
Mirantis this week made available a reference architecture for building and deploying artificial intelligence (AI) workloads using a control plane designed to be deployed on Kubernetes clusters.
Company CTO Shaun O’Meara said the Mirantis AI Factory Reference Architecture will make it simpler to build multi-tenant environments to run these applications using multiple classes of processors. The Mirantis AI Factory Reference Architecture is designed to enable IT teams to assemble infrastructure from reusable templates spanning compute, storage, networking and graphical processor units (GPUs) from NVIDIA, AMD and Intel.
Based on Mirantis k0rdent AI, an extension of an open source control plane that Mirantis launched earlier this year, it promises to reduce the amount of time needed to prototype, iterate, and deploy AI models and related services. It does so by providing, for example, access to curated integrations for tools for building applications, the continuous integration/continuous delivery (CI/CD) platforms used to deploy them and frameworks such as Gcore Everywhere Inference and NVIDIA AI Enterprise.
Mirantis via this layer of abstraction is also simplifying a range of additional complex issues, including providing remote direct memory access (RDMA) networking, allocation and slicing for GPUs, scheduling requirements, performance tuning and Kubernetes scaling.
Designed to support multiple types of AI workloads running on dedicated or shared servers; virtualized environments such as KubeVirt and OpenStack; public cloud or hybrid/multi-cloud; and edge computing locations, the Mirantis k0rdent AI centralizes the provisioning, configuration, and maintenance of AI infrastructure, including storage and networking services.
In general, IT teams are underestimating the need to isolate AI workloads running on shared infrastructure, said O’Meara. As more workloads are built and deployed in production environments, the need to isolate them across shared infrastructure resources is going to become significantly more pronounced simply because few organizations can afford to allocate IT infrastructure to a single application, he added.
Additionally, many organizations are also going to be required to navigate digital sovereignty requirements that stipulate workloads must run within a specific geographic region, noted O’Meara.
Unfortunately, many IT teams lack the IT infrastructure management expertise required to deploy AI workloads, an issue that Mirantis is hoping the Mirantis k0rdent AI control plane will help alleviate, he said.
In fact, many organizations are underestimating the total cost of AI by not factoring in the amount of IT infrastructure that will be required to run these types of workloads at scale, O’Meara added.
It’s still early days so far as deployment of AI workloads running at scale in production environments is concerned, but the one thing that is clear is that AI applications are not inexpensive to build and deploy. Nevertheless, as the pressure to operationalize AI continues to increase, the number of infrastructure challenges that IT teams will encounter will only continue to expand, especially as responsibility for managing the IT infrastructure needed to run these application shifts from data science to IT operations teams. The issue then becomes not just finding the best way to streamline the management of that IT infrastructure, but also justifying the level of investment that is going to be required to actually succeed.
