AI Computing &System Architecture
Generative AI PIM
3D-stacked PIM architecture, JEDEC-compliant, optimized for generative AI inference
This technology adopts a 3D-stacked Processing-in-Memory (PIM) architecture, vertically integrating logic and memory dies to bring computation closer to where data resides. By minimizing data movement, it effectively reduces latency and power consumption while overcoming the memory bandwidth limitations of the conventional von Neumann architecture. The system is JEDEC-compliant, ensuring compatibility with existing memory devices and System-on-Chip (SoC) platforms for seamless system integration.
The architecture integrates PIM computing cores with an efficient dataflow management mechanism to support Generative AI inference, accelerating key operations such as matrix multiplication and attention mechanism. This significantly improves inference performance and energy efficiency, making it well suited for edge AI, intelligent devices, and data center applications.
