AIR-LLM: Broadcasting AI Weights over Radio for Memory-Free Edge LLM Inference via RF Computing
Summary
The article proposes AIR-LLM, a memory-free edge LLM inference architecture that broadcasts AI weights over RF to edge devices. It leverages OFDM/MIMO to encode weights and uses edge RF mixers to perform GEMV in the RF domain, with edge-calibration units that compensate the channel through precoder/postcoder optimization; evaluated via Sionna ray-traced channels and real mixer profiling, showing orders-of-magnitude improvements in energy and airtime with multi-user broadcasting.