Accelerator Chaining for Efficient AI/ML Workloads in Kubernetes

Explore how to efficiently handle large AI/ML workloads in Kubernetes using accelerator chaining in this 37-minute conference talk from CNCF. Learn about extending Kubernetes with Custom Resource-based architecture and operators to orchestrate and configure device chains like FPGAs, GPUs, TPUs, and ASICs. Discover the benefits of direct data transfer between devices, including reduced memory copies, decreased CPU overheads, and lower latency. Gain insights into deploying these workloads easily and understand future developments with Dynamic Resource Allocation (DRA) support and CNI extensions. Presented by Sampath Priyankara from Nippon Telegraph and Telephone Corporation and Masataka Sonoda from Fujitsu Limited, this talk offers valuable knowledge for optimizing AI/ML performance in Kubernetes environments.