Open Infrastructure is not limited to Ironic or Cyborg nor Gyan in current forms. The eco system requires building and engaging with upstream AI models and libraries through Gyan as well Abstractions & Operators to be defined over Ironic, Cyborg and beyond. Hints to scheduler is just the Intent for placement. To map and optimize given workload, the Infrastructure resources like GPU, TPU, FPGA, SmartNIC, requires Domain specific modeling, resource Pooling, Defining, Designing & Deploying the workloads and closed loop models in production systems via standard clusters to optimize per cluster as well horizontally using multiple clusters. The goal of this presentation is to provide Best Known Configuration (BKC) for Cluster/Servers/PODs for workloads using Open Infrastructure Project called "Medhavi" and optimize through placement, image, and mappings whose requirements is still work-in-progress and the concept is to support the use cases such as image, voice & handwriting recognition requirements using "Medhavi" Object API.
Attendees will learn about Classifying workloads for AI/ML & 5G/Edge/IoT . What are the Models that encompass domain specific Knowledge, Relations and Attributes as Abstractions that are applicable for specific use cases and associated libraries as Operators. What virtual resources and device drivers the libraries and use case need to converge with constraints and error, precision etc. They will also learn how to map Abstractions and Operators frameworks to quantify through metrics the optimizations. Examples of use of Gyan, Medhavi, Ironic, Cyborg and related upstream resources.
