[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-82019-en":3,"doc-seo-82019-105":30,"detail-sidebar-cat-0-en-105":92},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":13,"seo_description":14,"update_tm":28,"read_time":29},82019,7971461740886,"Theodore","https://ap-avatar.wpscdn.com/davatar_3d24733baf745e90a7e4bdd5f77d97b2",8,"Research & Report","CRIMP Compact Reliable DNN Inference on In-Memory Processing via Crossbar-Aligned Compression and Non-ideality Adaptation","Crossbar-based in-memory processing (IMP) accelerators accelerate deep neural network inference but face incompatibilities with floating-point arithmetic, excessive hardware overhead from existing quantization, inefficient crossbar usage due to redundancy, and accuracy degradation caused by crossbar device non-idealities. CRIMP addresses these challenges with crossbar-friendly quantization using reusable bit-shift units, model compacting via kernel-group pruning and crossbar pruning to remove alignment units, and runtime-aware non-ideality adaptation integrated into one training framework for co-optimization and improved accuracy without hardware overhead.","arXiv :2607 .080 15v 1 [ cs .AR] 9 Jul 2026  \nCRIMP: Compact & Reliable DNN Inference on In-Memory Processing via Crossbar-Aligned Compression and  \nNon-ideality Adaptation  \nSHUO HUAI and HAO KONG, School of Computer Science and Engineering & HP-NTU Digital Manufacturing Corporate Lab, Nanyang Technological University, Singapore  \nXIANGZHONG LUO and SHIQING LI, School of Computer Science and Engineering, Nanyang Technological University, Singapore  \nRAVI SUBRAMANIAM, CHRISTIAN MAKAYA, and QIAN LIN, HP Inc., United States WEICHEN LIU∗ , School of Computer Science and Engineering, Nanyang Technological University, Singapore  \nCrossbar-based In-Memory Processing (IMP) accelerators have been widely adopted to achieve high-speed and low-power computing, alleviating the memory wall issues in the Von Neumann architecture, especially for the deep neural network (DNN) models with numerous parameters and high computational complexity. However, the floating-point (FP) arithmetic is not compatible with crossbar architectures. Although quantization schemes can remove FP parameters, current quantization techniques still require specific FP processors for multiplying scaling factors, incurring large hardware overhead. Besides, redundant parameters of current DNN models occupy too many crossbars, limiting the efficiency of crossbar accelerators. IMP-aware pruning methods are proposed to reduce the number of used crossbars, but current methods require data aligning among crossbars, which introduces significant memory overhead and computing overhead. On the other hand, due to the inherent non-ideal behavior of crossbar devices, like write variations, pre-trained DNN models suffer from accuracy degradation when it is deployed on a crossbar-based IMP accelerator for inference. Although some compensation methods are proposed to diminish the impact of device non-ideality, they introduce much overhead to the hardware design or runtime and do not consider the characteristic of crossbars. Especially, to deploy complex models on IMP accelerators, we should compact the model and mitigate the influence of device non-ideal behaviors without introducing significant overhead from each technology.  \nIn this paper, we first propose to reuse bit-shift units in crossbars for approximately multiplying scaling factors in our quantization scheme to avoid using FP processors. Second, we propose to apply kernel-group pruning and crossbar pruning to eliminate the hardware units for data aligning. Third, we adopt the runtimeaware non-ideality adaptation to relieve the impact of non-ideality from the training stage by exploiting the feature of crossbars. Finally, we integrate these three optimization procedures into one training process to forma comprehensive learning framework for co-optimization, which reduces the training overhead and achieves higher accuracy. The experimental results indicate that our quantization method incurs only a negligible  \n∗Corresponding author: Weichen Liu ([liu@ntu.edu.sg](liu@ntu.edu.sg))  \nsg, School of Computer Science and Engineering, com; Christian Makaya,  \n.  \nAuthors’ addresses: Shuo Huai, [shuo001@e.ntu.edu.sg](shuo001@e.ntu.edu.sg); Hao Kong, [kong.hao@ntu.edu.sg](kong.hao@ntu.edu.sg), School of Computer Science and  \nEngineering & HP-NTU Digital Manufacturing Corporate Lab, Nanyang Technological University, Singapore, Singapore;  \nXiangzhong Luo, [xiangzho001@e.ntu.edu.sg](xiangzho001@e.ntu.edu.sg); Shiqing Li, [shiqing.li@ntu.edu](shiqing.li@ntu.edu).  \nNanyang Technological University, Singapore, Singapore; Ravi Subramaniam, ravi.subramaniam@hp. [christian.makaya@hp.com](christian.makaya@hp.com); Qian Lin, [qian.lin@hp.com](qian.lin@hp.com), HP Inc., Palo Alto, California, United States; Weichen Liu, liu@ntu [edu.sg](edu.sg), School of Computer Science and Engineering, Nanyang Technological University, Singapore.  \nPermission to make digital or hard copies of all or part of this work for personal or classroom use is granted ","cbCaimgPTs6hWqU1","https://ap.wps.com/l/cbCaimgPTs6hWqU1","pdf",1700626,5,1,21,"English","en",105,"# 1 Introduction\n# CRIMP Framework Overview\n## Quantization with Bit-Shift Units\n## Kernel-Group and Crossbar Pruning\n## Runtime-Aware Non-ideality Adaptation\n# Experimental Results and Evaluation","[{\"question\":\"What problem does CRIMP target in crossbar-based in-memory processing for DNN inference?\",\"answer\":\"CRIMP targets floating-point incompatibility with crossbar architectures, overhead from quantization scaling-factor multiplication, inefficiency from redundant parameters occupying too many crossbars, and inference accuracy loss due to device non-idealities.\"},{\"question\":\"How does CRIMP avoid using floating-point processors in its quantization scheme?\",\"answer\":\"CRIMP reuses bit-shift units in crossbars to approximately multiply scaling factors, eliminating the need for floating-point processors.\"},{\"question\":\"How does CRIMP reduce hardware overhead caused by data alignment and device non-idealities?\",\"answer\":\"CRIMP applies kernel-group pruning and crossbar pruning to eliminate hardware units for data aligning, and uses runtime-aware non-ideality adaptation to relieve non-ideality effects during the training stage by exploiting crossbar characteristics.\"}]",1784177620,53,{"code":4,"msg":31,"data":32},"ok",{"site_id":25,"language":24,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":87,"head_meta":89,"extra_data":91,"updated_unix":28},"crimp-compact-reliable-dnn-inference-on-in-memory-processing-via-crossbar-aligned-compression-and-non-ideality-adaptation","",{"@graph":36,"@context":86},[37,54,69],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":50},"https://docshare.wps.com/document/research-report/",3,{"item":52,"name":13,"@type":43,"position":53},"https://docshare.wps.com/document/crimp-compact-reliable-dnn-inference-on-in-memory-processing-via-crossbar-aligned-compression-and-non-ideality-adaptation/82019/",4,{"url":52,"name":13,"@type":55,"author":56,"headline":13,"publisher":58,"fileFormat":61,"inLanguage":24,"description":14,"dateModified":62,"datePublished":63,"encodingFormat":61,"isAccessibleForFree":64,"interactionStatistic":65},"DigitalDocument",{"name":9,"@type":57},"Person",{"url":41,"name":59,"@type":60},"DocShare","Organization","application/pdf","2026-07-29","2026-07-16",true,{"@type":66,"interactionType":67,"userInteractionCount":20},"InteractionCounter",{"@type":68},"ViewAction",{"@type":70,"mainEntity":71},"FAQPage",[72,78,82],{"name":73,"@type":74,"acceptedAnswer":75},"What problem does CRIMP target in crossbar-based in-memory processing for DNN inference?","Question",{"text":76,"@type":77},"CRIMP targets floating-point incompatibility with crossbar architectures, overhead from quantization scaling-factor multiplication, inefficiency from redundant parameters occupying too many crossbars, and inference accuracy loss due to device non-idealities.","Answer",{"name":79,"@type":74,"acceptedAnswer":80},"How does CRIMP avoid using floating-point processors in its quantization scheme?",{"text":81,"@type":77},"CRIMP reuses bit-shift units in crossbars to approximately multiply scaling factors, eliminating the need for floating-point processors.",{"name":83,"@type":74,"acceptedAnswer":84},"How does CRIMP reduce hardware overhead caused by data alignment and device non-idealities?",{"text":85,"@type":77},"CRIMP applies kernel-group pruning and crossbar pruning to eliminate hardware units for data aligning, and uses runtime-aware non-ideality adaptation to relieve non-ideality effects during the training stage by exploiting crossbar characteristics.","https://schema.org",{"og:url":52,"og:type":88,"og:title":13,"og:site_name":59,"og:description":14},"article",{"robots":90,"canonical":52},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":93},[94,98,102,106,110,115,120,123,128,131,135],{"id":21,"doc_module":4,"doc_module_name":46,"category_name":95,"show_sort_weight":96,"slug":97},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":99,"show_sort_weight":100,"slug":101},"Literature",80,"literature",{"id":53,"doc_module":4,"doc_module_name":46,"category_name":103,"show_sort_weight":104,"slug":105},"Exam",70,"exam",{"id":20,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":46,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":46,"category_name":137,"show_sort_weight":20,"slug":138},19,"General","general"]