Modular General Purpose GPU for parallel computing to boost machine learning in avionics
The Avionic GPGPU+ project aims to research and develop a modular board based on a GPU (Graphics Processing Unit) core to power parallel computing and drive machine learning applications in avionics. This innovative concept was awarded funding by the current Ministry of Economic Affairs and Digital Transformation of Spain, under the Impulse Digital Enabling Technologies (THD) program.
The different technologies developed will allow for the completion of tests at the laboratory level, obtaining a proof of concept (TRL3), and reaching, at the end of the project, a functional prototype (TRL4) which will demonstrate its breakthrough differentiating elements, in computing power, performance and low consumption.
Artificial intelligence is currently becoming the industrial revolution of our era, providing the necessary intelligence for the extraction of knowledge and relevant information, which facilitates decision making in real time. The extraction of this knowledge, in order to improve decision making, depends on having large volumes of quality data, whether structured, unstructured or semi-structured, which needs to be processed by resource-intensive algorithms.
In CPU only based systems, intensive calculations consume a large amount of resources, even blocking tasks of the operating system itself, with high energy consumption, which does not make its intensive use viable. This is why parallel computing based on the use of GPUs (Graphic Processing Unit) is postulated as the roadmap for the paradigm shift towards distributed processing or Edge Computing, with increasingly powerful nodes or computers that allow data to be analyzed in a distributed manner, in real time, without the need to send massive amounts of data to centralized nodes.
Avionic GPGPU+ works in conjunction with a CPU, or even in parallel with more GPGPU+ cards depending on computational needs, providing direct access to the instruction set and cores in parallel to increase the computing performance of Clue Technologies’ WittyBox family of products.
With this objective, within the project Clue has investigated the state of the art, designed and verified the optimal architecture, and manufactured a functional prototype (TRL4), which has performed through the selected computer framework, executing performance comparisons to study and characterize its computational balance, its energy consumption and its optimization model, facilitating parallel computing and boosting machine learning applications in avionics, enabling direct application in the aircraft itself.
In conclusion, the Avionic GPGPU+ concept provides modularity and high performance in distributed computing, allowing, without the need of sending massive data to centralized nodes, while minimizing latency and with real time execution, to apply the HPC concept in Clue’s WittyBox family. This will ensure a compact form factor, low power consumption and competitive prices with breakthrough performance in our state of the art line of products.
The project has been developed at the Clue HQ development facilities in Málaga, following the company’s agile R&D development methodologies and structuring the project in 6 main Work Packages (WP) which provide specific milestones:
Firstly, in WP1, the activities are focused on the capture of requirements to develop the system specification, the state of the art of artificial intelligence algorithms for deep learning, and the preliminary architecture of the complete modular system, while in WP2, it will focus on the investigation of the camera module to be integrated, being necessary the purchase and procurement of different models to study its feasibility.
Once the architecture, high and low level requirements have been specified, in this fundamental package for the development of the project, the Avionic GPGPU+ concept will be designed, both mechanically and electronically, and the deep learning algorithm will also be researched and developed to allow its execution in low cost embedded systems, ensuring optimal behavior in operation, where the neural network will be trained with synthetic data. For this, it will be necessary to have the different development and server kits to perform the training and testing tasks.
Once the system has been designed, during the PT4 manufacturing phase it will be necessary to provide critical components to build and assemble the final prototype, where the previously designed neural network algorithm will be adjusted to optimize its behavior in the ad-hoc designed hardware. Finally, its performance will be validated and the relevant result reports will be generated in order to proof TRL4.
After the behavior of a unit has been verified, its group operation is validated, where the different nodes will have to analyze the images and gestures captured in real time, in order to send the data segmented by type of detection through BLE (Bluetooth Low Energy) wireless communications upon request.
WP1: REQUIREMENTS DEFINITION, ARCHITECTURE AND REGULATORY ANALYSIS
The first work package focuses on requirements capture to develop the system specification, define its functionalities and services, and determine the high and low level system requirements, in order to be able to propose the preliminary specifications of the connection interfaces, the state of the art of GPU based platforms, and the preliminary architecture of the complete system, within the regulatory framework for avionics.
WP2: GPGPU RESEARCH AND DEVELOPMENT
In this second work package, we will investigate the current state of the art to implement a GPGPU core for intensive computation, analyzing the technologies, their integration, reviewing that their specifications are compatible with the OpenVPX standard and with the main artificial intelligence frameworks.
WP3: CONCEPTUAL AND PRELIMINARY DESIGN (TRL3)
Once the architecture, the high and low level requirements are finalized, in this fundamental package for the development of the project, the GPGPU based electronic board compatible with OpenVPX will be designed, both mechanically and electronically, ensuring its compatibility with the modular computer systems.
WP4: SYSTEM AND MANUFACTURING REVIEW (TRL4)
Once the GPGPU modular board is designed, the integration of components at hardware level will be verified. Once the modular GPGPU card is designed, the integration of components at hardware level will be verified, its independent operation through monitoring software and finally it will be integrated in a modular computer to see its compatibility in the final system, validating its functional integration and generating the pertinent results reports (TRL4), prior to the performance tests of PT5.
WP5: PERFORMANCE AND OPTIMIZATION GPGPU CORE
Once the designed hardware is validated, it is necessary to quantify its performance, stressing the unit to measure its performance DVFS, ILP, TLP, DLP, and finally creating a model for the optimization of its performance, in order to be able to define the characteristics and potential of the implemented design.
WP6: PROJECT MANAGEMENT
This package is developed continuously during the execution of the project to ensure that the estimated deadlines are met in obtaining the planned objectives with the allocated budget and applying contingency methods for future deviations. Project management based on Agile methodology will be applied to streamline each design phase, minimize deviations and finally obtain a functional TRL4 prototype within the indicated timeframe.
The current importance of GPUs lies on the rebirth of DeepLearning, based on different neural network approaches, such as Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN), extensive in AI and cognitive computing. Deep learning powers many scenarios, including autonomous cars, disease diagnosis, computer vision, speech recognition and others, and is making great strides in the aeronautics sector. Within this framework, GPUs have become indispensable elements with the same importance that a CPU had years ago in any system. The integration of GPUs in different systems allows to bring to Edge Computing capabilities that were unimaginable, maintaining high throughput, low latency and ensuring data privacy.
The Avionic GPGPU+ project has been a success for Clue Technologies in this aspect, laying a very important technological foundation at TRL4 for the company and for the national aviation sector. The result, which will have the commercial name WittyGraphTM (WG), will be integrated as a modular card within the WittyBox family, which will increase the computing power of the WittyBox product family, and enhance Edge Computing solutions such as HUMS-DU (Health Monitoring System Data Unit) Predictive Maintenance.
The project has been executed with a high degree of efficiency and very satisfactory prototype results. It has suffered very limited deviations taking into account the impact of the SARS-CoV-2 virus and the supply crisis in the semiconductor sector worldwide that occurred during development. This high level of project control has occurred primarily thanks to Clue’s risk management methodologies
The following deliverables have been elaborated as part of the generated documentation according to the project plan:
- D1 Requirements definition, architecture and regulatory analysis.
- D2 Report on GPGPU technology for AI application enhancement.
- D3 Preliminary design review
- D4 TRL4 system review
- D5 Open VPX compatible prototype (Test Bench)
- D6 Open VPX compatible prototype (Module Integration)
- D7 TRL4 performance report Part 1
- D7 TRL4 performance report Part 2
- D8 Project planning
- D9 Results, dissemination and exploitation plan
- D10 Final project report
In conclusion, the development of the Avionic GPGPU+ project will establish a new line of work towards higher TRLs that will allow in the future for the incorporation of the architecture to the company’s main modular avionics equipment. This will allow for a plethora of new possibilities in applied use cases, as it becomes feasible to access in real time innovative AI solutions on-board aircraft systems such as:
- Optimization of resources (fuel) and maintenance operating costs.
- Proposal of more efficient routes
- Take-off time prediction
- Other aircraft predictive maintenance indicators.
- Complete aircraft autonomous management