通过计算与网络融合构建节能型人工智能基础设施.pdf

1、Towards Energy-Efficient AI Infrastructure through the Integration of Computing and NetworkingPreferred Networks,Inc.Towards Energy-Efficient AI Infrastructure through the Integration of Computing and NetworkingHirochika AsaiFellow/Preferred Networks,Inc.OPTICAL COMMUNICATION NETWORKSOutline(Optiona

2、l)54321Preferred NetworksPreferred Networks(PFN)develops key advanced AI technologies including AI chips,foundation models,and AI products.In collaboration with world-leading organizations,PFN leverages its accumulated deep expertise to integrate optimal solutions and drive innovation across a wide

3、range of industries.Scaling Law*:Performance of trained models is scaled by the following 3 parametersModel size(#of model parameters)Dataset size for trainingComputational resources(Peta-FLOPS-days)Massive computing resources used for AI196019701980199020002010202020302040地球Cray-1CDC-6600Source:htt

4、ps:/ AcceleratorTrend of#1 of TOP500Doubles in 24 months(Moores Law)Doubles in 3-4 monthsDoubles in 20 monthsTrend of HPCPerceptronNETtalkAVVINNRNN for SpeechLeNet-5BiLSTM for SpeechDeep Belief NetsAlexNetResNetsT17 Dota 1v1Neural MachineTranslationAlphaGoZeroPetaflop/s-daysTD-Gammon v2.11e-141e-121

5、e-101e-81e-61e-41e-21e-01e-+21e+41e+61e+81e+101e+12HPLFlopsDQNNWTFrontier富岳京Software/Algorithm optimizationFuture EraFirst EraDL Era1 Gflops1 Mflops1 Kflops1 flops0.001flops1 Tflops1 Zflops1 Eflops1 Pflops1 YflopsGPT-3PaLMTOP500 No.500TOP500 No.1PaLM2(*)J.Kaplan et al.,“Scaling Laws for Neural Langu

6、age Models,”arXiv:2001.08361Energy-efficiency is a key challenge in AI computing infrastructure!#$%&()*+),)-$+.()/0%12/2+.)+(2+2%*3(24)-)2+-3Energy-efficiency improvement of GPU(FP16)based on the peak performance/TDPMN-Core:XPU for AI ComputingGeneric Computing“Deterministic computing”MN-CoreMore re