Easy and affordable access to GPUs for AI/ML workloads
summary
Dynamic Workload Scheduler (DWS) allocates GPU/TPU capacity on demand for short ML runs by letting users request a number of GPUs and a duration, then queuing jobs when capacity appears; Flex Start gives just-in-time capacity, Calendar Mode guarantees a start time with short reservations. One customer, Two Sigma, integrates its scheduler with DWS-enabled GKE APIs in under two weeks to get hundreds of GPUs and Two Sigma boosted obtainability to >80% (from ~20%). Vertex Training already surfaces DWS and teams use Spot for cheaper inference.
AI-written summary · about
chapters (8)
- 0:00Meet Laura and Ari
- 1:04What is Dynamic Workload Scheduler?
- 3:21Which workloads function with Dynamic Workload Scheduler?
- 4:59How to choose between Compute Engine models
- 6:42Combining different Compute Engine models
- 8:32Real world examples
- 10:37Get started with Dynamic Workload Scheduler
- 11:20Wrap up
related talks
34mGPU as a Service Over K8s: Drive Productivity and Increase Utilization
21mUnlocking How To Efficiently, Flexibly, Manage and Schedule Seven AI Chips in Kubernetes
31mAccelerate AI adoption: automated cloud efficiency
20mHow are You Utilizing GPUs? Best Practices on Managing GPUs in Azure
09mDemocratizing Access to GPUs for Faster MLAI Innovation with Generative AI
38mMinimizing GPU Cost for Your Deep Learning on Kubernetes
31mShared GPU Scheduling & Proactive Autoscaling: A Production Blueprint for 1000+…
31mAI, CERN, and the Quest for GPU Custody: How CERN Leverages DRA for Efficient GPU Sharing