DC25_PAPER_Track1_ANovelThermoFlowUniformizerApplied_Son_V2.pdf

1、 Information Classification:General A Novel ThermoFlow Uniformizer Applied Immersion Cooling System(TFU-ICS)for Highly-Dense Multi-GPU-HBM Based Compute Module for AI Supercomputer Keeyoung Son,KAIST keeyoungkaist.ac.kr,+82-10-8506-7299 Taein Shin,KAIST Haeyeon Rachel Kim,KAIST Hyunjun An,KAIST Joun

2、gho Kim,KAIST Information Classification:General Abstract AI supercomputers need more GPUs and higher bandwidth for high AI computing performance,however it increasing power consumption causes thermal issues that degrade GPU performance and affect signal integrity(SI),reducing bandwidth.To meet the

3、demands of advanced AI models,higher bandwidth and computing density are crucial.Immersion cooling system(ICS)with high cooling capacities are emerging as thermal solutions,but future AI supercomputers need even better cooling solutions.With stable thermal reliability,SI performance can improve,allo

4、wing closer GPU integration,higher computing density,and lower power consumption.Therefore,we proposed novel thermoflow uniformizer applied ICS(TFU-ICS)adapated multi-GPU-HBM based compute module for future AI supercomputer.The proposed TFU is novel heatsink that control flow regime of immersion coo

5、lant for enhancing ICS performance and causing uniform thermal distribution of compute module.TFU is designed to adjust immersion coolant flow rate by varying its thickness,considering power profile of compute module,resulting in stronger and uniform performance.Based on powerful cooling performance

6、 of TFU-ICS,we designed much compact and highly-dense multi-GPU-HBM,with 4 GPUs integrated on interposer,not board.We also designed interconnection between each GPUs on same interposer and GPUs on other compute module.This compute module provides higher bandwidth with low power than conventional com