Persistent State Machines: LLM Attention with INT4 In-Memory Cells

Researchers present Persistent State Machines (PSMs), a formal framework for implementing LLM attention operators using INT4 in-memory cells with deterministic state transitions, backed by complete mathematical proofs for quantization error bounds and computational complexity. The approach was validated on FPGA hardware, achieving extremely low dynamic power consumption (below 1.0 mW for a 1024-cell array) and successful system-level integration on UltraScale+ devices with minimal resource utilization. The work demonstrates hardware feasibility for efficient, low-power LLM inference through deterministic discrete computation rather than traditional matrix operations.