Persistent State Machines: LLM Attention with INT4 In-Memory Cells
3 hours ago
- Persistent State Machines (PSMs) provide a formal discrete framework for attention operators in Large Language Models with complete mathematical proofs including quantization error bounds and DSPACE(O(n)) membership.
- Low-power implementation on a Zynq-7000 device achieves dynamic power below 1.0 mW and normalized dynamic energy of 3.81 × 10^-5 pJ/op using self-activation gating.
- System-on-Chip integration on an UltraScale+ device with PCIe Gen3 x1 bridge occupies only 0.67% of logic slices and achieves timing closure at 62.5 MHz.
- Functional simulation over 1000 random vectors confirms bit-exact agreement with a fixed-point software reference, while all energy figures are simulation-based and exclude physical board measurement.