利用硅光子技术开发人工智能加速器 - Sang Yoon Han（DGIST）.pdf

1、Han 1Development of AI AcceleratorsLeveragingSilicon Photonics TechnologySangyoon Han,DGIST2025/June/15thHan 2Silicon Photonics AI Transfer Learning EngineSangyoon Han,DGIST2025/June/15thHan 3HAETAE project overviewjoongang.co.kr Haetae Beast in Korean folklore Cow+horse+lion with horn A creature de

2、feating fireHan 4Goal:Multi-Material-Chiplet Photonic Transfer Learning Engine Toward the realization of optical computings ultimate vision By Chiplets(Si photonics)Heterogeneous integration(Multi-material)Near-zero static power MEMSHan 5The team 2 from Korea,3 from EUGreeceBelgiumGermanyKoreaKoreaP

3、rof.Gunther Roelkens Prof.Kyoungsik YuDr.Leonardo Del BinoProf.Nikolaos PlerosProf.Sangyoon HanHan 6Optical computingHan 7Calculation using electrons Complexity emerges from simplicityhttp:/hyperphysics.phy-astr.gsu.edu/hbase/Electronic/nand.htmlhttps:/ 8Calculation using light Simple Han 9Calculati

4、on using light1023201020.Han 10Calculation using light =4 2x4 Lens1011122 Han 11Matrix computation using lens.10135.571.28.31.45411.=.Han 12Matrix computation using lensX.Lin et al.,Science 2018Han 13Paradigm Shift:Computing with Light Fundamental breakthrough possiblePhysics-based optimization Leas

5、t-Time,Least-Energy Nature-Driven OptimizationParallelismSpatial,modal,and wavelength divisionMulti-dimensional paradigmUltra-BroadbandCarrier frequency:100 THzWavelength multiplexing:Dissipationless Lossless transmission of information Even better at higher bandwidthLow latency Calculation through

6、propagation of light Latency=distance/cQuantum nature of light Intrinsic Robustness to Thermal FluctuationsMcMahon,P.L.The physics of optical computing.Nat Rev Phys 5,717734(2023).Han 14Leveraging Silicon PhotonicsHan 15Photonic Integrated Circuits(PICs)“Photonics on a chip”Integrated with electroni