AMD filed 2 patents this quarter across Graphics (1) and AI & Machine Learning (1).
The Graphics patent describes a ray tracing architecture that caches BVH topology separately from geometry data to accelerate scene updates in dynamic real-time environments. In AI & Machine Learning, AMD filed a hierarchical reinforcement learning system where leader NPCs train follower NPCs through demonstrations, aiming to produce more efficient and realistic behaviors in games and virtual environments.
AMD's Graphics patent tackles one of the most computationally expensive aspects of real-time ray tracing by separating BVH topology from geometry data. Traditional ray tracing rebuilds the entire bounding volume hierarchy when objects move, but this approach caches the tree structure that defines spatial relationships while allowing geometry positions to update independently. The topology can then be reused across frames even as objects shift in dynamic game environments, reducing the processing overhead that typically limits ray tracing performance.
The AI & Machine Learning patent describes a hierarchical training system where experienced NPCs demonstrate behaviors to less experienced ones, creating a teacher-student relationship within the game world itself. Instead of training every NPC individually from scratch, follower NPCs learn by observing leader NPCs, reducing computational requirements while still producing diverse behaviors across different skill levels and roles. This demonstration-based approach aims to create more adaptive AI that can handle varying complexities without proportionally increasing resource demands.