人工智能普及的人生之旅.pdf

1、The Journey to AI PervasivenessVictor PengPresident,AMD2|1960s1989Early History:Supervised Learning,Perceptrons,MNISTEvolution of AIV11 M Chip1984Early NPUs at ISSCC1991Alpha 2126419983|1960sEarly History:Supervised Learning,Perceptrons,MNISTEvolution of AI19892012201420162018202020242022Transformer

2、sChat GPTAgentic AIMulti-Model MOEs SSMsDeep Neural Networks enabled by ImageNet and GPUs AlexNetFirst FPGA SOCAMD Zynq 70402013First AIE DeviceVC19022021Largest FPGA EverVP19022022Leadership GPU for AI InferenceMI300X2024Charlie Giattino,Edouard Mathieu,Veronika Samborska and Max Roser(2023)-“Artif

3、icial Intelligence”Published online at OurWorldInData.org.Retrieved from:https:/ourworldindata.org/artificial-intelligence Online Resource4|Todays AIThe Journey to AI Pervasiveness in SupercyclesAI PervasivenessLarge-Scale Trainingand InferenceCloudPersonal Assistants,Games,CopilotEndpointsAutomotiv

4、e,Sensor Intelligence,Robotics,MedicalEmbeddedVirtual Personalized PhysicalWith Increasing Unit Volume5|Todays AIThe Investment and Value Creation RaceAI PervasivenessReported revenue for 2023,internally estimated numbers based on https:/ visualization?20232024(est)2025(est)CAPEX DC AIData Center AI

5、CAPEX Data Center AI Data Center AI Revenue6|Todays AIThe Challenges AheadAI PervasivenessAdditional challenges:High quality data,safety,reliability,confidentialityOn the technology side:Foremost power source&distribution,and diverse requirements of diverse use cases7|Power to Train Frontier ModelsB

6、ased on published parameter counts of leading training models;and AMD internal calculations approximating energy to trainGPT-3 175BGPT-4oEst.10T modelEst.100T model1.00E+001.00E+011.00E+021.00E+031.00E+041.00E+051.00E+061.00E+071.00E+081.00E+091E+091E+101E+111E+121E+131E+141E+15MWhours#ParametersEne