Adaptive DTC Liquid Cooling for AI Clusters_ Managing Variable Workloads and Improving Energy Efficiency.pdf

1、Adaptive DTC Liquid Cooling for AIClusters:ManagingVariableWorkloadsandImprovingEnergyEfficiencyMontse VilarrubCOO&Co-Founder-UniSCoolDavid Beberide Business Development Director-UniSCoolAI CLUSTERSThermal Bottlenecks in AI ClustersWhen static cooling meets dynamic workloadsAI scaling is constrained

2、 by thermal limits,making advanced cooling essential for next-generation AI clusters.3%worldwide electrical power 2%worldwide GHG emissions 40%due to cooling Limitations of todays DTCWhen static cooling meets dynamic workloadsTransient and non-uniform CPU/GPU workloads cause bottlenecks,lower perfor

3、mance,and faster degradation.UniSCools solutionAdaptive cooling,just where its neededAdaptive heat sink that locally deform to adapt the thermal resistance as a function of the local temperature.Developed by+10 years research500k invested U.S.Patented technologyUniSCools solutionGoals to be achieved

4、Energy savingsHotspot mitigation&high temperature uniformity without IHSIncreased chip lifetime and reliability Temperature(C)GoalSmart CoolMicrochannelsUniSCools solutionTechnology validationBenchmark setup:Smart Cool vs.microchannels in a CDU-based cooling system on a Dell R760 server.Realistic co

5、nditions:AI workloads with spatial and temporal heat variations,mimicking real AI clusters.Key capability tested:Temperature uniformity and energy savings under dynamic and non-uniform heat flux.Validation:Experimental demonstration of improved thermal homogeneity under AI-driven load conditions Uni

6、SCools solution69%temperature uniformity improvement.Up to 825C improvement in temperature uniformity compared to microchannels(depending on load)UniSCools solution69%hydraulic power energy savings The impact of Adaptative CoolingSmarter heat extraction,measurable resultsEnergy s