© 2019 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes,creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. “The final publication is available at: DOI: 10.1109/DSD.2019.00064 GPU4S: Embedded GPUs in Space Leonidas Kosmidis∗,Jer´ omeˆ Lachaizey, Jaume Abella∗ Olivier Notebaerty, Francisco J. Cazorla∗;z, David Steenarix ∗Barcelona Supercomputing Center (BSC), Spain yAirbus Defence and Space, France zSpanish National Research Council (IIIA-CSIC), Spain xEuropean Space Agency, The Netherlands Abstract—Following the same trend of automotive and avion- in space [1][2]. Those studies concluded that although their ics, the space domain is witnessing an increase in the on-board energy efficiency is high, their power consumption is an order computing performance demands. This raise in performance of magnitude higher than the limited power budget of a space needs comes from both control and payload parts of the space- craft and calls for advanced electronics able to provide high system, which is limited to a couple of Watts. computational power under the constraints of the harsh space Interestingly, GPUs entered in the embedded domain to environment. On the non-technical side, for strategic reasons it is satisfy the increasing demand for multimedia-based hand- mandatory to get European independence on the used computing held and consumer devices such as smartphones, in-vehicle technology. In this project, which is still in its early phases, we entertainment systems, televisions, set-top boxes etc. They study the applicability of embedded GPUs in space, which have shown a dramatic improvement of their performance per-watt were re-designed compared to their high-performance desk- ratio coming from their proliferation in consumer markets based top/server counterparts to exhibit low-power requirements, on competitive European technology. To that end, we perform essential for battery power and thermally-constrained devices. an analysis of the existing space application domains to identify Improvements in the transistor technology allowed achieving which software domains can benefit from their use. Moreover, impressive performance capabilities that were only possible we survey the embedded GPU domain in order to assess whether embedded GPUs can provide the required computational power in high performance systems of the past decade [3]. Although and identify the challenges which need to be addressed for their the GPU vs CPU performance ratio is lower in the embedded adoption in space. In this paper, we describe the steps to be domain than in the high-performance one due to power and followed in the project, as well as the results of our preliminary thermal constraints, mobile GPUs are increasingly considered analyses in the first months of the project. for accelerating heavy workloads, for applications ranging from signal processing, to advanced driving assistance systems I. INTRODUCTION AND BACKGROUND (ADAS) in cars, as well as prototype supercomputers for The space market is in constant search for high performance, exascale [4][5]. scalable processing solutions to satisfy the increased compu- Despite their promising characteristics, embedded GPUs tational needs of future missions for increased autonomy and have not been explored for their applicability in this domain. data processing. While reusing solutions from other domains This project aims at covering this gap by providing an initial can reduce non-recurrent costs, space has its unique set of assessment of existing embedded GPUs, as a first step of their constraints. Increased performance demands are required for further exploration and adoption in this domain. the platform computers, which are in charge of controlling The rest of the paper is organised as follows. Section II critical functionalities like the spacecraft power distribution, introduces the project, which includes the on-going and future navigation and guidance, as well as for the payload comput- activities. Section III analyses the performance demand of ers (in charge of controlling the payload devices, and pre- current and future space missions. Section IV details the processing acquired payload data before its transmission to outcome of work on the analysis of the embedded GPU the ground) to process more data. market. Finally, Section V describes the future work and Graphics processing units (GPUs), initially a special pur- Section VI presents the main conclusions of our work so far. pose type of accelerator for visualisation tasks, have since sev- eral years ago outperformed Central Processing Units (CPUs) II. THE PROJECT raw performance and energy efficiency. This opens the door to The project explores the suitability of embedded GPUs for achieve unprecedented performance with a very high energy space from both, software and hardware perspectives. In order efficiency for demanding computations, becoming essential to reach its main goal, the project is organised in 4 main for high-performance computing. As a matter of fact, one activities as shown in Figure 1. quarter (125) of the supercomputers in the recent edition of Space mission analysis: In this activity, we survey potential the TOP500 / Green500 list (as of June 2019) are based on space areas and particular algorithms and applications that GPUs, including the two most powerful supercomputers. Past can benefit from the use of GPUs in space, ensuring that studies analysed the applicability of high-performance GPUs the identified applications are suitable for the GPU execution forms. As a next step we will perform a thorough comparison Space mission analysis Candidate Next steps for platforms GPU adoption between the selected embedded GPU candidates and existing Embedded GPU analysis evaluation in Space and future processing devices for the space domain. This way, the potential benefits of embedded GPUs can be evaluated, April 2019 time specifically regarding the high-performance requirements of future missions, while respecting the power and thermal lim- Fig. 1. Main activities. Box width does not represent activity duration itations of the space environment. In order to do this, we will perform the evaluation by porting representative kernels derived from existing space algorithms to the GPU. Moreover, model. This is important because there are fundamental design performance and other data from past missions and previous differences between GPUs and CPUs which are required to Application-Specific Integrated Circuit (ASIC) processes will be taken into account to understand how certain algorithms be used for comparison, by normalising them according to used in space, e.g. those with divergent path execution among the current space technology node (65nm). This will allow different threads or those not performing coalesced memory the selection of the most appropriate GPU platform for space accesses, behave differently in embedded GPUs w.r.t. high- from the evaluated candidates. performance GPUs. In particular, certain space algorithms Next Steps for GPU adoption: As final step of the project, may not be suitable for embedded GPUs despite their high we will define the future steps required towards the adoption computational nature, and might require to be redesigned. of GPUs in space, by identifying current limitations and We provide a preliminary list of applications already vali- proposing appropriate solutions to overcome them. In order dated as GPU-compatible in Section III. Moreover, we define to assess the next steps for the adoption of GPUs in the a preliminary set of criteria that can be used for the selection space domain with a system integrator view, we will examine of a GPU candidates depending on the mission profile. the necessary steps for qualification of COTS systems by Embedded GPU analysis: In parallel to previous activity, addressing their reliability concerns at system level or the we survey the available commercial-of-the-shelf (COTS) hard- development of radiation hardened components. In the former ware and soft-IP (Intellectual Property), in order to identify case we will also provide software fault-tolerance solutions their characteristics with the goal of selecting a candidate specifically designed for these platforms, in case a COTS board for evaluation in the next activities. We mainly focus embedded GPU is selected in the previous task. Finally, an on European IP, which in turn dominates the market and can ESA provided algorithm will be ported to the GPU platform. provide complete independence in the European space sector. In this line, in order to understand the wide range of embedded III. SPACE APPLICATION SURVEY GPUs, we perform a taxonomy of existing products. In this Section, we examine the space applications domain In addition to the mission profile criteria from the Space regarding the use of embedded GPUs. The performance re- mission analysis, an important aspect for the selection of quirements of space missions are constantly increasing. As an a candidate board, is the performance and energy-efficiency example of current missions, the Gaia astrometry mission characteristics, due to the fundamental architectural differ- (devoted to the measurement of the positions of celestial ences between their designs and that of their high-performance bodies), launched in 2013 by the European