Meta has unveiled Andromeda, a proprietary machine learning system designed to enhance ad recommendation through advanced retrieval capabilities. The system utilizes NVIDIA Grace Hopper Superchip and Meta Training and Inference Accelerator (MTIA) hardware.
Key Performance Improvements
System Enhancements
- 6% recall improvement in retrieval systems
- 8% ads quality improvement on selected segments
- 3x enhancement in end-to-end model inference queries per second
- 10,000x increase in model capacity
Advantage+ Creative Impact
- 22% ROAS increase for new Advantage+ creative users
- 7% conversion increase with image generation
- Over 15 million ads created monthly using GenAI tools
- 1 million+ advertisers utilizing GenAI capabilities
Technical Innovations
Deep Neural Network Architecture
- Custom-designed for NVIDIA Grace Hopper Superchip
- Sublinear inference cost
- Enhanced personalization capabilities
- GPU preprocessing for feature extraction
Hierarchical Indexing System
- Multi-layered ad organization
- Reduced inference steps
- Joint training of index and retrieval models
- Improved precision and recall versus traditional methods
Model Optimization
- Segment-aware design
- Real-time complexity adjustment
- 10x boost in model inference efficiency
- Advanced kernel fusion techniques
Future Development
The system will expand to include:
- Autoregressive loss function integration
- MTIA hardware compatibility
- Projected 1,000x increase in model complexity
- Enhanced ad diversity capabilities
This advancement represents Meta's commitment to improving advertiser outcomes while managing the exponential growth in creative content through AI-driven automation.
