AIHacker Newsabout 3 hours ago
Persistent State Machines Cut LLM Attention Power
1 min read
Persistent State Machines formalize LLM attention with INT4 in-memory cells. A 1024-cell Zynq-7000 array estimates dynamic power below 1.0 mW. A 256-cell UltraScale+ SoC closes timing at 62.5 MHz, occupying 0.67% of logic slices.
Level
Hype check
Tap to vote and see what everyone thinks.
Summary by ByteBrief
