Keynote – Aparna Ramani
summary
Meta describes a deep redesign of data infra to serve ML at scale including rack designs that link thousands of GPUs, new liquid cooling and faster networks, and co‑designed hardware software stacks. Empirical details include 175 billion parameter model, training on 940 million Instagram samples versus 14 million ImageNet samples yielding a ~5% accuracy gain, 80% IO wasted before feature flattening, and pipelines that ingest tens of billions of samples and retrain ranking models every few hours.
AI-written summary · about
chapters (4)
- 0:00Start
- 8:55Data Infra must be re-designed beyond just incremental changes
- 13:35Data freshness has dramatic impact on model quality
- 14:34ML adds exponential complexity to observability
related talks
02h 16mKeynote Day 2 - Full
18mThe AI-First Data Infrastructure
52mKeynote Interview
28mAccelerating AI at Scale with Data
22mFireside Chat: Evolution of AI-first Data Infrastructure
17mKeynote Analysis
28mApplied ML at Scale Strategies & Insights
13mCombining algorithms and quality data to build better AI platforms