Pipeline Parallel vs Tensor Parallel Two Clusters of 2 x 24GB M4 Pro Mac Mini ($1,399 ea). Both running unquantized Llama 3.1 8B (fp16). Tensor Parallel uses low-latency RDMA over Thunderbolt. 15.7 tok/sec vs 28.4 tok/sec (1.8x). Close to linear scaling. (EXO Labs)