[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84352-en":3,"doc-seo-84352-105":29,"detail-sidebar-cat-0-en-105":90},{"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},84352,1099514068365,"Aurelia","https://ap-avatar.wpscdn.com/avatar/10000253d8d9f28188e?_k=1776742907772140068",8,"Research & Report","Empirical Analysis of GPU Frequency Behavior Under ML Workloads","This work investigates NVIDIA GPUs’ frequency-scaling behavior under ML/AI workloads through ongoing empirical research. Preliminary results show that on lower-performance and thermally constrained GPUs, operating frequency is strongly shaped by recent workload history, typically within an ~80 ms window. This undermines latency-prediction approaches that assume per-kernel independence and sum isolated measurements, because dynamic frequency scaling creates inter-kernel dependencies. Future directions include improved latency models, kernel reordering, and frequency/latency/energy-aware design guidance.","Empirical Analysis of GPU Frequency Behavior  \nUnder ML Workloads  \nTruong-Thanh Le†, Hoang-Loc La§ , Amir Taherkordi†, Frank Eliassen†, Phuong Hoai Ha§ and Peiyuan Guan†  \n† Department of Informatics  \nUniversity of Oslo Oslo, Norway {truongl,amirhost,frank,[peiyuang](peiyuang}@ifi.uio.no)[}](peiyuang}@ifi.uio.no)[@ifi.uio.no](peiyuang}@ifi.uio.no)  \n§ Department of Computer Science UiT The Arctic University of Norway  \nTromsø, Norway {[hoang.l.la](hoang.l.la),[phuong.hoai.ha](phuong.hoai.ha}@uit.no)[}](phuong.hoai.ha}@uit.no)[@uit.no](phuong.hoai.ha}@uit.no)  \narXiv :2607 .08307v 1 [ cs .DC] 9 Jul 2026  \nAbstract—This work presents ongoing research on the frequency-scaling behavior of NVIDIA GPUs when executing ML/AI workloads. Our preliminary findings show that, on lower-performance GPUs, the operating frequency is strongly affected by the recent workload history—typically within an 80ms window. This behavior challenges a common assumption underlying several state-of-the-art ML latency-prediction techniques, which treat individual GPU kernel latencies as independent and therefore estimate total execution time by summing isolated per-kernel measurements. Our results indicate that such an assumption does not always hold, as the GPU’s dynamic frequency scaling introduces inter-kernel dependencies. We also outline several promising directions for leveraging this observation in future work, including improved latency-prediction models, GPU kernel-reordering strategies, and NAS-driven guidelines for frequency/latency/energy-aware model design.  \nIndex Terms—NVIDIA, GPU Frequency, Kernel Throughput, Performance Modeling  \nI. INTRODUCTION  \nThese days, modern GPUs span a wide range of performance levels, from compact mobile chips to large data-center accelerators. Due to these differences in thermal design, power delivery, and efficiency requirements, the runtime behavior of such devices can vary significantly, leading to distinct performance patterns—especially in how they manage operating frequency under ML and AI workloads. Hence, understanding how the dynamic frequency-scaling mechanisms behave on these different classes of GPUs is crucial for accurately characterizing their performance under ML and AI workloads.  \nWhen running ML/AI workloads, the GPU often scales its operating frequency under the control of an internal management core, which dynamically adapts the frequency based on recent or ongoing activity. This controller typically considers factors such as the number of arithmetic operations executed, the volume of data transferred across the DRAM–L1/L2 cache hierarchy, thermal condition, power limit, and a variety of hidden factors [3] . Unfortunately, understanding the precise decision-making policies of this controller is difficult, as both its firmware and the GPU-kernel implementations of modern ML/AI models are closed-source, leaving these mechanisms largely opaque to researchers.  \nDue to this complexity, most state-of-the-art methods for predicting the performance of ML/AI workloads on modern  \nGPUs simply ignore such behavior, assuming instead that each kernel operates independently. For example, Paleo [5], Habitat [2], NeuSight [4], and PM2Lat [3] all adopt a similar strategy for estimating model latency on specific devices: they categorize GPU kernels into multiple groups, assign a predictive model or parameter set to each category, and then estimate the total execution time by summing the predicted latency of individual kernels. Although several approaches attempt to capture end-to-end behavior by treating a model asa sequence of kernels and applying sequence-based learning techniques—such as graph neural networks [6], or LSTMs [1]—to predict total latency, these methods still fall short of generalizing across different GPU architectures and model types, making them less efficient and limiting their ability to accurately capture real performance variations in practice.  \nOur ongoing work examines GPU frequenc","cbCaivosA1yvCzbC","https://ap.wps.com/l/cbCaivosA1yvCzbC","pdf",1677773,1,4,"English","en",105,"# I. Introduction\n# II. Experiments\n## A. Kernel-level Frequency/Throughput analysis","[{\"question\":\"What is the main focus of the empirical study?\",\"answer\":\"The study analyzes how NVIDIA GPUs scale their operating frequency when running ML/AI workloads, especially how frequency responds to workload history.\"},{\"question\":\"What key finding challenges existing ML latency-prediction assumptions?\",\"answer\":\"On lower-performance GPUs, frequency depends strongly on recent workload history, so treating kernel latencies as independent (and summing them) can be inaccurate due to inter-kernel dependencies.\"},{\"question\":\"How is the GPU frequency updated over time according to the findings?\",\"answer\":\"For laptop-class and thermally constrained GPUs, the controller uses an approximate 80 ms window of preceding activity to choose the frequency for the next 20 ms interval.\"}]",1784195003,10,{"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":85,"head_meta":87,"extra_data":89,"updated_unix":27},"empirical-analysis-of-gpu-frequency-behavior-under-ml-workloads","",{"@graph":35,"@context":84},[36,52,67],{"@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":21},"https://docshare.wps.com/document/empirical-analysis-of-gpu-frequency-behavior-under-ml-workloads/84352/",{"url":51,"name":13,"@type":53,"author":54,"headline":13,"publisher":56,"fileFormat":59,"inLanguage":23,"description":14,"dateModified":60,"datePublished":61,"encodingFormat":59,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":55},"Person",{"url":40,"name":57,"@type":58},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":20},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"What is the main focus of the empirical study?","Question",{"text":74,"@type":75},"The study analyzes how NVIDIA GPUs scale their operating frequency when running ML/AI workloads, especially how frequency responds to workload history.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"What key finding challenges existing ML latency-prediction assumptions?",{"text":79,"@type":75},"On lower-performance GPUs, frequency depends strongly on recent workload history, so treating kernel latencies as independent (and summing them) can be inaccurate due to inter-kernel dependencies.",{"name":81,"@type":72,"acceptedAnswer":82},"How is the GPU frequency updated over time according to the findings?",{"text":83,"@type":75},"For laptop-class and thermally constrained GPUs, the controller uses an approximate 80 ms window of preceding activity to choose the frequency for the next 20 ms interval.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":57,"og:description":14},"article",{"robots":88,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,109,114,119,122,127,130,133],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":21,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":110,"doc_module":4,"doc_module_name":45,"category_name":111,"show_sort_weight":112,"slug":113},6,"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":28,"doc_module":4,"doc_module_name":45,"category_name":131,"show_sort_weight":28,"slug":132},"Lifestyle","lifestyle",{"id":134,"doc_module":4,"doc_module_name":45,"category_name":135,"show_sort_weight":105,"slug":136},19,"General","general"]