用于神经形态计算技术的硅兼容氧化铪基铁电体 - Athanatos Dimoulas.pdf

1、Vision Transformers with Ferroelectric OxidesCoordinatorETHZM.DelucaSALA.DimoulasNCSRDT.KaempfeFraunhoferYang-Ho GongSoongsil UKyeong-Sik MinKookmin UMin Hyuk ParkSeoul UD.KwonHanyang UL.Begon-LoursBridging Europe and South Korea toBuild the Next Generation of AI Chipshttps:/vitfox.eu/12ndROK-EU for

2、um,16thJune 2025,Jeju,ROKSi-compatible Hafnia-based ferroelectrics for neuromorphic computing technologiesVision Transformers-Specifications2Transformers:From natural language processing(e.g.GPT)to computer vision-Vision transformers replace Convolutional Neural Networks(CNN)for image recognition-Vi

3、T outperform CNN with 4x fewer computing resources.ViTFOX project objectives Design and fabrication of CIM demonstrator implementing a single layer network,comprising of:sense amplifiers,logic controller,32x32 epi-FTJ synaptic weights,32x32 1T-4C 3D FeRAM arrays Energy efficiency 30 TOPS/W,accuracy

4、90%Design circuit level simulator integrating:compact models of ferroelectric oxides,CMOS and peripherals energy efficiency 50 TOPS/Watt HW-SW co-optimization for ViT with ferroelectric oxides 30%energy improvement over DRAM2ndROK-EU forum,16thJune 2025,Jeju,ROKViTFOX strengthens the leading positio

5、n of EU and KoreaHafnia-based,Si-compatible ferroelectric electronics Pioneered in Europe(Germany)and significantly advanced by Korean researchers,members of this project.32ndROK-EU forum,16thJune 2025,Jeju,ROKStabilization of the ferroelectric phase:doping(Si,Zr,.),stress,annealing,film thickness D

6、iscovered in 2007-first published 2011,QimondaCompatible with Si processing T.S.Bscke et al,Appl.Phys.Lett.99,102903(2011)50%Zr H0.5Zr0.5O2(HZO)Fluorite Hafnia-based Ferroelectric Oxides42ndROK-EU forum,16thJune 2025,Jeju,ROK5Ferroelectric NVM“flavors”Versatility:Functionality,integration,applicatio