IBM and ETH Zrich Researchers Unveil Analog Foundation Models to Tackle Noise in In-Memory AI Hardware

IBM researchers in collaboration with ETH Zrich have developed a new class of Analog Foundation Models (AFMs) that aim to integrate large language models (LLMs) with Analog In-Memory Computing (AIMC) hardware, addressing the longstanding challenge of noise-induced errors in AIMC systems. AIMC offers significant efficiency advantages by performing matrix-vector multiplications directly within dense non-volatile memory (NVM) arrays, eliminating the von Neumann bottleneck and enabling high throughput and low power consumption, which is crucial for deploying AI models on edge and embedded devices. The primary obstacle for AIMC adoption has been