Computing Large Foundational Models
summary
A stability.ai engineering overview that states concrete scale and outcomes: a cluster of 4,100 A100 GPUs on AWS with a secondary Weave cluster and H100 coming for 3–10× performance gains; image latency drops from ~30 seconds for 512×512 to a few seconds via better samplers, compilation and GPUs; models shipped include Stable Diffusion, Stable Diffusion XL (≈3× larger, 1024-native with text encoder), DeepFloyd (cascaded diffusion with a 4.6B first stage), APIs, DreamStudio, upscaler, Animation API, voice conversion demos and a LoRA/adapter fine-tuning pipeline.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 2:01Exploring Stability AI
- 8:01Partnerships and Integrations
- 12:01API and Platform Features
- 20:01Animation API and Creative Examples
- 24:01Controllability and Future Applications
- 26:31Conclusion and Future Plans
related talks
29mFrom Experiment to Enterprise: Architecting AI for Stability and Scale
13mFrom Silicon to Scale: Training at Scale
48m5 Steps to Deploy Cloud Native Sustainable Foundation AI Models
17mCharacterization of Large Language Model Development in the Datacenter
31mScaling AI for Industry
20mTraining a foundation image model and deploying it to millions of users
27mFoundation Models: Blueprints for success
54mThe Future of AI: Building the Most Impactful Technology Together