Falcon: An Event-Driven Object Tracking SoC with Activation-Skip and Weight-Compression MRAM PIM and Heterogenous Q/K/V Read-Only-Once Attention Flow

This paper presents Falcon, an event-driven object tracking SoC integrating heterogeneous MRAM computing blocks with three features: 1) An activation-skip and weightcompression MRAM processing-in-memory (PIM) for CNN encoders with 2.44 × efficiency and 61.9% latency reduction. 2) A Q/K/V read-only-o