{"id":{"repo_id":"athens","oai_identifier":"oai:lib.uoa.gr:uoadl:2962089"},"canonical_url":"https://search.ndltd.org/etd/athens/oai:lib.uoa.gr:uoadl:2962089","repository":{"repo_id":"athens","name":"National and Kapodistrian University of Athens","base_url":"https://pergamos.lib.uoa.gr/uoa/dl/frontend/oaipmh"},"display":{"title":"Dissecting the Performance of AI Applications Using NVIDIA GPUs at the Edge","abstract":"The vast amount of data stemming from an increasingly large number of end devices, such as User Equipment (UE) and Internet of Things (IoT), poses excessive bandwidth requirements and dramatically increased latency, when this data needs to be transferred to and processed by remote cloud sites. The advent of edge computing along with Artificial Intelligence, also abbreviated as edge intelligence (EI), attempts to solve this problem, allowing edge computing hardware closely located to the data sources to be combined with edge AI applications, such as AI inference. This thesis evaluates the performance of an object detection AI application on two NVIDIA Jetson edge computing devices. Specifically, both AI training and inference phases are evaluated against different settings that can impact application performance. The main contribution of this thesis is to provide experimental evidence on performance metrics (e.g., throughput) and power consumption, while assessing whether the NVIDIA Jetson devices can effectively conduct both AI training and inference. Specifically, the performance of the NVIDIA Jetson devices using different (i) processing engines, such as general-purpose Central Processing Units (CPUs) and (Accelerated) Graphics Processing Units (GPUs), (ii) power configurations, (iii) network architecture types, and (iv) values of training parameters is investigated. Our experimental results show that although the NVIDIA Jetson devices under test have enough capacity to perform both AI training and inference for the selected application, AI training is time-consuming on such low-end devices. On the other hand, AI inference on Jetson’s GPU-based processing engines achieves substantially higher performance than in the case of a CPU processing engine. Finally, apart from performance aspects, this thesis also addresses power consumption aspects of the tested processing engines (i.e., CPU and GPU).","abstract_html":"The vast amount of data stemming from an increasingly large number of end devices, such as User Equipment (UE) and Internet of Things (IoT), poses excessive bandwidth requirements and dramatically increased latency, when this data needs to be transferred to and processed by remote cloud sites. The advent of edge computing along with Artificial Intelligence, also abbreviated as edge intelligence (EI), attempts to solve this problem, allowing edge computing hardware closely located to the data sources to be combined with edge AI applications, such as AI inference. This thesis evaluates the performance of an object detection AI application on two NVIDIA Jetson edge computing devices. Specifically, both AI training and inference phases are evaluated against different settings that can impact application performance. The main contribution of this thesis is to provide experimental evidence on performance metrics (e.g., throughput) and power consumption, while assessing whether the NVIDIA Jetson devices can effectively conduct both AI training and inference. Specifically, the performance of the NVIDIA Jetson devices using different (i) processing engines, such as general-purpose Central Processing Units (CPUs) and (Accelerated) Graphics Processing Units (GPUs), (ii) power configurations, (iii) network architecture types, and (iv) values of training parameters is investigated. Our experimental results show that although the NVIDIA Jetson devices under test have enough capacity to perform both AI training and inference for the selected application, AI training is time-consuming on such low-end devices. On the other hand, AI inference on Jetson’s GPU-based processing engines achieves substantially higher performance than in the case of a CPU processing engine. Finally, apart from performance aspects, this thesis also addresses power consumption aspects of the tested processing engines (i.e., CPU and GPU).","abstract_has_math":false,"creators":["Srivallapanondh Sasipim"],"institution":null,"degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2021,"date_issued":"2021","date_published":"2021","updated_at":"2026-11-19T05:43:51Z","subjects":["Τεχνολογία – Πληροφορική","Technology - Computer science"],"languages":["English"],"rights":[],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier","label":"Identifier","values":["uoadl:2962089"],"render_values":[{"text":"uoadl:2962089","href":null,"code":true}]}],"person_identifier_entries":[]},"links":{"outbound_url":"https://pergamos.lib.uoa.gr/uoa/dl/object/2962089","outbound_label":"Repository record","outbound_source":"dc:identifier","record_url":"https://pergamos.lib.uoa.gr/uoa/dl/object/2962089","record_label":"Repository record","record_kind":"record","record_source":"dc:identifier","download_url":null,"download_label":null,"download_source":null,"access_status":null,"embargo_until":null,"access_note":null},"source_record":{"url":"https://pergamos.lib.uoa.gr/uoa/dl/frontend/oaipmh?verb=GetRecord&metadataPrefix=oai_dc&identifier=oai%3Alib.uoa.gr%3Auoadl%3A2962089","prefix":"oai_dc"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:creator","label":"Author","values":["Srivallapanondh Sasipim"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2021"]},{"key":"dc:type","label":"Dc Type","values":["born_digital_postgraduate_thesis","Διπλωματική Εργασία","Postgraduate Thesis"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Τεχνολογία – Πληροφορική","Technology - Computer science"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["English"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["uoadl:2962089","https://pergamos.lib.uoa.gr/uoa/dl/object/2962089"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["The vast amount of data stemming from an increasingly large number of end devices, such as User Equipment (UE) and Internet of Things (IoT), poses excessive bandwidth requirements and dramatically increased latency, when this data needs to be transferred to and processed by remote cloud sites. The advent of edge computing along with Artificial Intelligence, also abbreviated as edge intelligence (EI), attempts to solve this problem, allowing edge computing hardware closely located to the data sources to be combined with edge AI applications, such as AI inference. This thesis evaluates the performance of an object detection AI application on two NVIDIA Jetson edge computing devices. Specifically, both AI training and inference phases are evaluated against different settings that can impact application performance. The main contribution of this thesis is to provide experimental evidence on performance metrics (e.g., throughput) and power consumption, while assessing whether the NVIDIA Jetson devices can effectively conduct both AI training and inference. Specifically, the performance of the NVIDIA Jetson devices using different (i) processing engines, such as general-purpose Central Processing Units (CPUs) and (Accelerated) Graphics Processing Units (GPUs), (ii) power configurations, (iii) network architecture types, and (iv) values of training parameters is investigated. Our experimental results show that although the NVIDIA Jetson devices under test have enough capacity to perform both AI training and inference for the selected application, AI training is time-consuming on such low-end devices. On the other hand, AI inference on Jetson’s GPU-based processing engines achieves substantially higher performance than in the case of a CPU processing engine. Finally, apart from performance aspects, this thesis also addresses power consumption aspects of the tested processing engines (i.e., CPU and GPU)."]},{"key":"dc:title","label":"Title","values":["Dissecting the Performance of AI Applications Using NVIDIA GPUs at the Edge"]}]}],"canonical_facts":{"dc:creator":["Srivallapanondh Sasipim"],"dc:date":["2021"],"dc:description":["The vast amount of data stemming from an increasingly large number of end devices, such as User Equipment (UE) and Internet of Things (IoT), poses excessive bandwidth requirements and dramatically increased latency, when this data needs to be transferred to and processed by remote cloud sites. The advent of edge computing along with Artificial Intelligence, also abbreviated as edge intelligence (EI), attempts to solve this problem, allowing edge computing hardware closely located to the data sources to be combined with edge AI applications, such as AI inference. This thesis evaluates the performance of an object detection AI application on two NVIDIA Jetson edge computing devices. Specifically, both AI training and inference phases are evaluated against different settings that can impact application performance. The main contribution of this thesis is to provide experimental evidence on performance metrics (e.g., throughput) and power consumption, while assessing whether the NVIDIA Jetson devices can effectively conduct both AI training and inference. Specifically, the performance of the NVIDIA Jetson devices using different (i) processing engines, such as general-purpose Central Processing Units (CPUs) and (Accelerated) Graphics Processing Units (GPUs), (ii) power configurations, (iii) network architecture types, and (iv) values of training parameters is investigated. Our experimental results show that although the NVIDIA Jetson devices under test have enough capacity to perform both AI training and inference for the selected application, AI training is time-consuming on such low-end devices. On the other hand, AI inference on Jetson’s GPU-based processing engines achieves substantially higher performance than in the case of a CPU processing engine. Finally, apart from performance aspects, this thesis also addresses power consumption aspects of the tested processing engines (i.e., CPU and GPU)."],"dc:identifier":["uoadl:2962089","https://pergamos.lib.uoa.gr/uoa/dl/object/2962089"],"dc:language":["English"],"dc:subject":["Τεχνολογία – Πληροφορική","Technology - Computer science"],"dc:title":["Dissecting the Performance of AI Applications Using NVIDIA GPUs at the Edge"],"dc:type":["born_digital_postgraduate_thesis","Διπλωματική Εργασία","Postgraduate Thesis"]},"updated_at":"2026-11-19T05:43:51Z"}