[Opening Keynote] Making Distributed Computing Easy
summary
AI workloads outpace single machines — compute demand grows 35x every 18 months while per‑core performance inches up, so training and serving go distributed. Examples: online learning pipelines stitch Spark, Kafka, Flink for ingestion, Horovod/distributed TensorFlow for training and Seldon/TensorFlow Serving for inference; reinforcement learning needs large‑scale simulations (Horizon, DeepMind/OpenAI buildouts). Ray provides a universal framework with an in‑memory object store, and Anyscale adds automatic file sync, env‑var no‑code cloud runs, serverless autoscaling, Docker image builds, SDK/REST and dashboards to simplify deploy.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 0:42Distributed applications are becoming the norm
- 3:05Building distributed applications is hard
- 7:50Two basic approaches
- 8:57Ray
- 10:02Development Production
- 12:29AnyScale
related talks
49mDistributed Deep Learning
03mRay: A Distributed Execution Framework for Emerging AI Applications
12mLightning Talk: Distributed AI Without the Infrastructure Tax
06mSponsored Keynote: The Weight of Data: Rethinking Cloud-Native Systems for the Age of AI
15mRay: A Distributed Compute Engine for AI
25mOpening Keynote
34mPanel Discussion: Training and Inference at Planet Scale
26mUsing Ray On Large-scale Applications at Ant Group