UNIPOOL: Unlocking Communication Potential of LLM System via Interface Pooling Architecture and Schedule

Abstract

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.

Publication
Proceedings of the 33rd IEEE International Symposium on High-Performance Computer Architecture (HPCA 2027)
Shuwen Deng
Shuwen Deng
Associate Professor