AI/ML Infra Meetup | RAYvolution - The Last Mile: Mastering AI Deployment with Ray
0 likes264 views
The document discusses three major trends in AI infrastructure, focusing on scaling laws, AI data processing, and heterogeneity. It emphasizes the need for larger models, more compute, and data to achieve better results, while also highlighting the transition from SQL-centric to AI-centric data processing. Additionally, it addresses the challenges posed by multimodal data and diverse hardware requirements in modern AI applications.
AI/ML Infra Meetup | RAYvolution - The Last Mile: Mastering AI Deployment with Ray
- 1. Three Developments in AI Infra Robert Nishihara Co-creator of Ray, Co-founder of Anyscale
- 2. My Background 2013: Began PhD in AI 2016: Started Ray along with Philipp Moritz and Ion Stoica 2019: Started Anyscale to commercialize Ray
- 6. Three Industry Trends How the world has changed since we started Ray
- 8. Three Industry Trends Scaling Laws ● Larger models, more compute, more data → Better results ● Top priority for many AI infra teams: “enable training on 100x more data” ● ML platform / infra teams are on the critical path ● Scaling data -> scaling compute
- 9. Three Industry Trends Scaling Laws ● Larger models, more compute, more data → Better results ● Top priority for many AI infra teams: “enable training on 100x more data” ● ML platform / infra teams are on the critical path ● Scaling data -> scaling compute Question: How can I use more compute to get higher quality data?
- 10. Three Industry Trends Scaling Laws ● Larger models, more compute, more data → Better results ● Top priority for many AI infra teams: “enable training on 100x more data” ● ML platform / infra teams are on the critical path ● Scaling data -> scaling compute AI Data Processing ● Data processing: SQL centric → AI centric ● Multimodal data (text, images, video, audio, …) ● New requirements: data intensive and GPU intensive
- 11. Three Industry Trends Scaling Laws ● Larger models, more compute, more data → Better results ● Top priority for many AI infra teams: “enable training on 100x more data” ● ML platform / infra teams are on the critical path ● Scaling data -> scaling compute AI Data Processing ● Data processing: SQL centric → AI centric ● Multimodal data (text, images, video, audio, …) ● New requirements: data intensive and GPU intensive Question: How can I use AI to get more insights from my data?
- 12. Three Industry Trends Scaling Laws ● Larger models, more compute, more data → Better results ● Top priority for many AI infra teams: “enable training on 100x more data” ● ML platform / infra teams are on the critical path ● Scaling data -> scaling compute AI Data Processing ● Data processing: SQL centric → AI centric ● Multimodal data (text, images, video, audio, …) ● New requirements: data intensive and GPU intensive Heterogeneity ● Multimodality, hardware accelerators, frameworks, clouds, models
- 13. Three Industry Trends Scaling Laws ● Larger models, more compute, more data → Better results ● Top priority for many AI infra teams: “enable training on 100x more data” ● ML platform / infra teams are on the critical path ● Scaling data -> scaling compute AI Data Processing ● Data processing: SQL centric → AI centric ● Multimodal data (text, images, video, audio, …) ● New requirements: data intensive and GPU intensive Heterogeneity ● Multimodality, hardware accelerators, frameworks, clouds, models