LLM workloads have different memory and communication bandwidth needs. UniPool combines processor-level interface pooling with runtime scheduling so systems can allocate bandwidth between memory and communication as workloads change.