Preferred Networks Announces MN-Core L2000 AI Accelerator
Japanese startup Preferred Networks (PFN) announced its collaboration with SEMIFIVE and Samsung to develop the MN-Core L2000 AI inference accelerator, targeting more than double the memory bandwidth of its predecessor. This project represents a crucial step in advancing AI inference speeds and workload energy efficiency through vertical stacking and advanced DRAM integration. It highlights the growing reliance on specialized custom silicon and foundry partnerships to bypass traditional 2D semiconductor scaling limitations. The MN-Core L2000 series is scheduled to complete its chip design by the first quarter of 2027, following PFN's roadmap for its 2027 product releases. It aims to improve upon the first-generation L1000's architecture, which featured 32GB of on-chip memory with over 10TB/s of bandwidth.
## BACKGROUND
Chip stacking technology involves vertically integrating multiple semiconductor dies to increase integration density and performance, helping overcome the slowdown of traditional two-dimensional dimensional scaling. SEMIFIVE operates as a custom semiconductor design house partnering with major foundries like Samsung to build application-specific integrated circuits.