Sokovan: An Acceleration-first Kubernetes-compatible AI Container Engine
summary
Sokovan is an acceleration‑first, Kubernetes‑compatible container engine that runs alongside K8s via a daemonset shim and custom operators, mapping logical resource groups to node pools and keeping direct hardware control for accelerators. It claims a 2022 milestone of 120+ Gb/s per multi-node GPU workload and 150+ Gb/s RDMA, implements fractional GPU scaling, supports >10 accelerator types, deploys in air‑gapped sites, and uses GDS and vendor‑aware storage acceleration; GA release targets early next year.
AI-written summary · about
chapters (15)
- 0:00Sokovan introduction
- 1:27AI infra challenges
- 3:12Kubernetes basics
- 6:43Why Kubernetes struggles with AI
- 9:54AI workload characteristics
- 12:04Sokovan’s design evolution
- 14:13AI stack deployment complexity
- 17:47Embrace Kubernetes ecosystem
- 20:42Sokovan’s Kubernetes integration
- 23:10Architecture on Kubernetes
- 25:45Disaggregated serving future
- 27:31Confidential computing AI agents
- 29:14Hosting no-code AI tools
- 30:20Enterprise deployment challenges
- 32:09Summary open source
related talks
38mBeyond Kubernetes: Adapting To Spe...
21mUnlocking How To Efficiently, Flexibly, Manage and Schedule Seven AI Chips in Kubernetes
30mRunning AI Workloads in Containers and Kubernetes
29mThe Path to GPU as a Service in Kubernetes
31mAccelerate AI adoption: automated cloud efficiency
25mMaking the Most Out of Your Hardware Accelerators in a Kubernetes Cluster
25mExtending Kubernetes for AI | Lessons Learned From Platform Engineering
03mThe Next Evolution of Kubernetes: GPU-Centric Infrastructure for AI Workloads