Preprint tests multi-PC pipeline for models beyond one machine's memory

Researchers reported a preprint describing a pipeline-parallel system that splits an LLM into pre-compiled OpenVINO shards across Intel AI PCs. In tests, a two-node Llama 3.1 8B INT4 pipeline served two concurrent users at 1.79 times the single-user throughput of unsplit inference on the same hardware. The authors also report that four Lunar Lake AI PCs on Intel Tiber Cloud served a 70B model that no single fleet member could hold at interactive speed.