In-memory computing slashes MCTS energy to 60 mW for Go AI

A new multi-primitive in-memory computing (IMC) strategy for Monte Carlo tree search (MCTS) has been introduced by researchers, promoting energy-efficient AI decision-making on edge devices. While traditional processors typically draw 55-300 W, the innovative IMC-MCTS method operates at approximately 60 mW for 9x9 Go, resulting in an energy efficiency increase of 96 times compared to a CPU and between 65 and 2,059 times against an H100 GPU. This method employs phase-to-primitive decomposition, transforming each algorithmic phase—selection, expansion, rollout, and backpropagation—into hardware-native IMC primitives, including content-addressable memory and a resistive random-access memory (RRAM) crossbar. Operating at 22 nm with RRAM-array parameters, the system maintains the entire search on-chip and achieves a European Go Federation rating within the uncertainty range of the original MCTS, challenging previous beliefs about IMC's compatibility with complex multi-phase algorithms and paving the way for new edge AI applications.