JExplore: Design Space Exploration Tool for Nvidia Jetson Boards
Basar Kutukcu 1 , Sinan Xie 1 , Sabur Baidya 2 , Sujit Dey 1 · 2025
Read JExplore: Design Space Exploration Tool for Nvidia Jetson Boards by Basar Kutukcu 1 , Sinan Xie 1 , Sabur Baidya 2 , Sujit Dey 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-15773.txt.
Introduction
Basar Kutukcu and co-authors present in a 2025 study JExplore, a tool for searching the best settings of Nvidia Jetson boards that run artificial intelligence at the edge. The boards expose many hardware knobs, and the best combination depends on the user's needs. Hand-tuning across boards is slow and inconsistent.
JExplore wraps any search algorithm in a shared test bench, handling the host-board chatter, configuration swaps, and metric logging. Researchers can plug in their own search method and compare fairly against others. The tool aims to turn a scattered craft into a common benchmarking ground for edge computing.
Catalog record
| CREATOR | Basar Kutukcu 1 , Sinan Xie 1 , Sabur Baidya 2 , Sujit Dey 1 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 14 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-15773
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
JExplore: Design Space Exploration Tool for Nvidia Jetson Boards Basar Kutukcu 1 , Sinan Xie 1 , Sabur Baidya 2 , Sujit Dey 1 Abstract Nvidia Jetson boards are powerful systems for executing artificial intelligence workloads in edge and mobile environments due to their effective GPU hardware and widely supported software stack. In addition to these benefits, Nvidia Jetson boards provide large configurability by giving the user the choice to modify many hardware parameters. This large space of configurability creates the need of searching the optimal configurations based on the user’s requirements. In this work, we propose JExplore, a multi-board software and hardware design space exploration tool. JExplore can be integrated with any search tool, hence creating a common benchmarking ground for the search algorithms. Moreover, it accelerates the exploration of user application and Nvidia Jetson configurations for researchers and engineers by encapsulating host-client communication, configuration management, and metric measurement. I Introduction The recent advancements of generative artificial intelligence have been affecting many different areas such as engineering, art, writing, chat for learning, etc. Many industries in these areas may benefit from running generative artificial intelligence models on private and mobile systems. Robotics, automotive, and consumer electronics are examples of such industries. There are many mobile hardware systems that can run generative artificial intelligence workloads such as mobile phones [ 1 ] , embedded systems with GPU support [ 2 ] and FPGAs [ 3 ] . However, these usually require lots of modifications to the models developed in common deep learning libraries such as PyTorch and HuggingFace. Nvidia GPUs are dominant hardware for deep learning workloads due to their computation power and widely adopted software stack. They have the advantage of being the most common hardware backend for many deep learning tools. This usually translates to ease of using different systems as long as they both have capable Nvidia GPUs without leaving the development ecosystem. This creates the first advantage of the Nvidia Jetson boards. Any model developed in the common deep learning libraries is directly usable in Jetson boards as they have the same Nvidia GPU architectures as the other Nvidia GPUs. The second advantage is that, even though small in size, Jetson boards carry quite powerful GPUs. Considering large language models, the workhorse of generative artificial intelligence, require large computation powers, a powerful hardware backend is a requirement as much as it is an advantage. Another advantage of Jetson boards is their configurability. Jetson boards allow users to change various hardware parameters such as number of CPU cores, CPU frequency, GPU frequency, EMC frequency etc. Modifying these parameters naturally results in changes in various performance metrics, such as application latency and power consumption. Finding close to optimal configurations in this vast search space requires intelligent search algorithms. These algorithms need to explore the search space by sampling configurations on the system. Since these explorations are hard to implement and take long time, many of the search algorithms are tested on exhaustively collected small search space [ 4 ] or synthetic problems [ 5 ] . However, using a small search space cannot test the scalability of the search algorithm and using a synthetic problem cannot test the algorithms reliability in real applications. Therefore, there’s a need for large, real world application based search spaces and a tool to interface this search space and search algorithms. Based on the advantages of Jetson boards and the necessity of an exploration tool for the search algorithms, we propose JExplore. The main contributions of this tool and the paper are: • Encapsulating communication, configuration management, metric measurement and result saving into simple and yet effective APIs • It can communicate with multiple Jetson boards, allowing batch sampling algorithms to work faster • Creating a common benchmarking ground for the search algorithms for testing in a large and real world application based search space • We have published the code as open source at https://github.com/basarkutukcu/JExplore II Motivation and Related Work Improvements in technology and a large variety of requirements from different applications result in large design spaces in many different platforms. As a result, search algorithms become important in identifying optimal solutions in these large design spaces. Many different search algorithms are proposed to search these design spaces. In [ 6 ] , Bayesian optimization with expected hypervolume improvement based acquisition function is used. In [ 7 ] , an evolutionary algorithm is used to solve large search spaces. In [ 2 , 8 ] , customized Bayesian optimization algorithms are used to search design spaces. While there are many algorithms like these, there are not many opportunities to compare these algorithms. One common methodology is to use synthetic problems such as [ 5 ] to benchmark and compare algorithms. While synthetic problems can create lots of different conditions and are very helpful to quickly benchmark a synthetic problem, they are not the real target of these search algorithms. Search algorithms are developed to solve real world problems and therefore testing against real world problems are important. There are also many different real world problems that create design spaces. In [ 8 ] , FPGA design problems are used to benchmark the search algorithm. In [ 4 ] , an FPGA design space, an ASIC design space and a compiler optimization design space are used. In [ 2 ] , combined search spaces of Nvidia Jetson hardware and various kinds of software parameters are used. In [ 9 ] , compiler optimization search space is used. The improvements in generative artificial intelligence are promising in terms of the future of generative AI based applications. A large portion of these applications can be on embedded systems to extend their benefits deep into people’s lives since embedded systems are ubiquitous in life. Since Nvidia GPUs are one of best hardware that can execute these workloads, exploring the design spaces posed by Nvidia Jetson is an important research problem. Therefore, our tool can accelerate the development of the search algorithms and provide a common useful benchmarking ground for the search algorithms. III JExplore Fig. 1: JExplore overview The overall architecture of JExplore is shown in Figure 1 . JExplore consists of two main components which are JHost and JClient. JHost runs on the host computer and works as an interface between the user defined search algorithm and Jetson boards. It also has utility functions such as saving the explored search space in CSV format. JHost and JClients on Jetson boards communicate using ZMQ Push/Pull sockets over SSH. IP addresses and ports are defined by the user when the instances of these classes are initialized. JClient runs on the Jetson device. It has 3 main capabilities. The first one is to configure the Jetson device and the application based on the input it receives from JHost. There are many modifiable hardware parameters in the Jetson systems, which make them flexible embedded systems. The full list of these parameters, the number of values they can take and the ranges of these values are shown in Table I . JClient can configure these parameters using its subcomponent called JConfig. Nvidia normally gives power modes to the user to modify these parameters. However, there are usually 5-10 modes provided