A user built a 40-node cluster using Intel N150 mini PCs with 12 GB DDR5 each, interconnected via 10 Gbps USB 3.2 in a tree topology. Running GLM 5.3 Flash at native precision, the cluster achieves 14 tokens per second with capacity for 8x1M context windows. The builder notes that 1 Gbps LAN interconnects were insufficient for this workload.
Read original
reddit/r/LocalLLM