Keynote: Programming the Cloud as Easily as Your Laptop
summary
Claims that training compute grows ~35x every 18 months and model size grew ~20x (200x in the last two years), while single‑core speed and even GPUs/TPUs cannot keep up, so distribution is mandatory. Presents a functioning Anyscale service built on Ray that auto‑syncs laptop code to the cloud, auto‑scales and allocates GPUs/TPUs, supports stateful training and serving, is multi‑cloud and serverless‑like, and exposes a dashboard for debugging. Demonstration-ready first release shows the infinite laptop illusion and a true serverless experience without typical limits.
AI-written summary · about
chapters (23)
- 0:00Introduction
- 0:15History of distributed computing
- 1:57Today distributed computing is hard
- 2:35What is different this time
- 4:10Performance of a single processor
- 4:27Single core performance
- 4:53Specialized hardware
- 5:41Memory performance
- 6:28AI centric applications
- 6:50Distributed workloads
- 7:24AI and HPC overlap
- 8:13Summary
- 8:53Today is extremely hard
- 9:21Three solutions
- 11:40What do we need
- 12:00What developers want
- 12:25Infinite laptop
- 12:56How we do this
- 13:16Properties of Anyscale
- 13:36Automatic Code Sync
- 14:10Serverless Experience
- 15:02Cloud Native
- 15:33Conclusion
related talks
17mKeynote
15mKeynote: From Platforms to AI Factories. Has Kubernetes Solved It?
59mKeynote: AI without the BS, for humans
22mKeynote
08mKeynote: How WebAssembly can Power the New Wave of Serverless AI
39mServerless at the end of the Universe
26mCanva's AI Leap: Scaling with Ray and Anyscale
28mThe Brave New World of serverless computing