ThAME: 3D Memory-Enabled Heterogeneous Accelerator for LLM Mixture of Experts
ThAME is proposed, a three-dimensional (3D) heterogeneous multi-chiplet architecture for MoE inference that employs Ferroelectric Field-Effect Transistor-based non-volatile and DRAM-based volatile memory chiplets with a co-designed compute mapping strategy that aligns the distinct computational profiles of attention me...