[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84823-en":3,"doc-seo-84823-105":29,"detail-sidebar-cat-0-en-105":91},{"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":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},84823,2336464648322,"Aria","https://ap-avatar.wpscdn.com/avatar/2200025388227c56fec?_k=1778556882303663488",8,"Research & Report","Performance Optimization and Comparative Analysis of Generative AI Models on Advanced Accelerators","Generative AI models such as LLMs and diffusion systems face practical deployment barriers, including large memory footprints, long inference latency, heavy compute requirements, and high hardware costs. Evaluation across heterogeneous platforms is further complicated by mismatches in numeric formats, memory bandwidth, and software stacks interacting with model architecture and workloads. The study presents a systematic performance optimization and comparative analysis across diverse downstream tasks, introducing mixed-precision post-training quantization evaluation, fine-tuning strategies, and measurements on modern HPC systems and advanced accelerators.","arXiv :2607 .05400v 1 [ cs .PF] 5 May 2026  \nPerformance Optimization and Comparative Analysis of Generative AI Modelson Advanced Accelerators  \nAMITASH NANDA∗ , University of California San Diego, USA  \nJAVIER HERNANDEZ NICOLAU, San Diego Supercomputer Center, USAMADHUSUDAN GUJRAL, San Diego Supercomputer Center, USAMAHIDHAR TATINENI, San Diego Supercomputer Center, USA AMITAVA MAJUMDAR, San Diego Supercomputer Center, USA DEBASHIS SAHOO, University of California San Diego, USA  \nGenerative AI models, such as Large Language Models (LLMs) and diffusion models, have demonstrated impressive performance across a wide range of tasks. Despite these advances, deployment remains challenging due to substantial memory requirements, extended inference latency, significant computational demands, and high hardware costs. These issues are further complicated when evaluating models across heterogeneous platforms, where differences in numerical formats, memory bandwidths, and software stacks interact with model architecture and workload characteristics in complex ways. To address these challenges, we present a systematic study focused on performance optimization and comparative analysis of several Generative AI models across diverse downstream tasks. This work introduces a novel mixed-precision post-training quantization evaluation, examines fine-tuning strategies, and assesses performance across modern high-performance computing (HPC) systems and advanced accelerators.  \nAdditional Key Words and Phrases: Optimization, Nvidia GPU, Intel Gaudi, Accelerators, Benchmarking, Quantization, LLMs  \n1 Introduction  \nIn recent years, generative AI models have been widely adopted by researchers, engineers, students, and others. Most are used for natural language processing, such as large language models (LLMs), or for image generation, such as diffusion models [2, 3, 6] . Both approaches are based on traditional neural networks and transformer architecture, and have been heavily optimized, with core operations (mainly matrix multiplications) that can be parallelized and offloaded to specialized accelerators. These accelerators are available in a wide variety of architectures from leading vendors and differ significantly between generations. In this work, we present our analysis and evaluation using accelerators from three different high-performance computing (HPC) systems: Perlmutter [8], Expanse [10], and Voyager [1] . Perlmutter is housed at NERSC, Lawrence Berkeley National Laboratory, while Expanse and Voyager are hosted at the San Diego Supercomputer Center (SDSC) . Perlmutter is a heterogeneous system comprising both CPU-and GPU-accelerated nodes. It consists of 1, 536 GPU nodes, each with 1 AMD Milan processor and 4 NVIDIA A100 GPUs. Expanse is also a general-purpose, heterogeneous distributed compute cluster, organized into 13 Scalable Compute Units (SSCUs), with 728 standard AMD CPU nodes, 54 NVIDIA V100 GPU nodes, 8 NVIDIA A100 GPU nodes, 34 NVIDIA H100 GPU nodes, and 4 large-memory nodes. Voyager, on the other hand, is an innovative system specifically designed for AI applications, comprising 42 nodes with Intel Gaudi accelerators. Currently, second-generation accelerators (Gaudi2) are replacing the original first-generation (Gaudi1) nodes. In addition, we were granted limited access to the third-generation Gaudi  \n∗ This author contributed majorly to this research.  \nAuthors’ Contact Information: Amitash Nanda, [ananda@ucsd.edu](ananda@ucsd.edu), University of California San Diego, San Diego, CA, USA; Javier Hernandez Nicolau, [jhernandeznicolau@sdsc.edu](jhernandeznicolau@sdsc.edu), San Diego Supercomputer Center, San Diego, CA, USA; Madhusudan Gujral, [mgujral@ucsd.edu](mgujral@ucsd.edu), San Diego Supercomputer Center, San Diego, CA, USA; Mahidhar Tatineni, [mahidhar@sdsc.edu](mahidhar@sdsc.edu), San Diego Supercomputer Center, San Diego, CA, USA; Amitava Majumdar, [majumdar@sdsc.edu](majumdar@sdsc.edu), San Diego Supercomputer Center, ","cbCaioaVQxdXAM9L","https://ap.wps.com/l/cbCaioaVQxdXAM9L","pdf",4935067,1,6,"English","en",105,"# Introduction\n## Mixed-Precision Post Training Quantization on Voyager and Perlmutter","[{\"question\":\"为什么生成式AI模型在真实硬件上部署仍然具有挑战？\",\"answer\":\"部署面临显著内存需求、较长推理延迟、较高计算开销和较高硬件成本；在不同平台上评估还会受到数值格式、内存带宽与软件栈差异的影响。\"},{\"question\":\"文中提出的“敏感性知晓”的混合精度PTQ框架做了什么？\",\"answer\":\"提出一种敏感性知晓的混合精度后训练量化框架，并采用两阶段方法在不同加速器上优化LLM模型性能。\"},{\"question\":\"作者在哪些任务与平台上进行了评估？\",\"answer\":\"以TinyLlama-1.1B模型为例，在语言建模（WikiText-2）、常识推理（HellaSwag）和阅读理解（BoolQ）三类任务上评估，并覆盖NVIDIA A100与Intel Gaudi等加速器。\"}]",1784198549,15,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":27},"performance-optimization-and-comparative-analysis-of-generative-ai-models-on-advanced-accelerators","",{"@graph":35,"@context":85},[36,53,68],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/performance-optimization-and-comparative-analysis-of-generative-ai-models-on-advanced-accelerators/84823/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"为什么生成式AI模型在真实硬件上部署仍然具有挑战？","Question",{"text":75,"@type":76},"部署面临显著内存需求、较长推理延迟、较高计算开销和较高硬件成本；在不同平台上评估还会受到数值格式、内存带宽与软件栈差异的影响。","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"文中提出的“敏感性知晓”的混合精度PTQ框架做了什么？",{"text":80,"@type":76},"提出一种敏感性知晓的混合精度后训练量化框架，并采用两阶段方法在不同加速器上优化LLM模型性能。",{"name":82,"@type":73,"acceptedAnswer":83},"作者在哪些任务与平台上进行了评估？",{"text":84,"@type":76},"以TinyLlama-1.1B模型为例，在语言建模（WikiText-2）、常识推理（HellaSwag）和阅读理解（BoolQ）三类任务上评估，并覆盖NVIDIA A100与Intel Gaudi等加速器。","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,114,119,122,127,130,134],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":21,"doc_module":4,"doc_module_name":45,"category_name":111,"show_sort_weight":112,"slug":113},"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":45,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":123,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":45,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":45,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":106,"slug":137},19,"General","general"]