Deep learning neural networks are a special subfield of AI that will play a key role in enabling fully autonomous driving. To drive a car without human intervention, however, it requires a complex sensor framework that captures not only vehicle data but also data from its surroundings. These sensors include LiDAR, radar, video, cameras, and more, which continuously generate massive amounts of data about the environment around the car in real time. Deep learning neural networks help synthesize data and create meaningful information that enables vehicles to react in real time.

    To date, most deep learning algorithms have been executed in clouds or data centers with powerful processors or GPUs with extensive cooling. This is not possible with self-driving cars. Due to the real-time requirements, it is necessary to perform target detection and other operations in the vehicle itself, and sending data to the cloud will not work properly. Therefore, self-driving cars also require specialized hardware to implement deep learning algorithms and meet performance, power, and cost requirements for massive production.

    Implementing deep learning algorithms in hardware is a challenge in itself. For example, a common object detection algorithm, which is based on a CNN (Convolutional Neural Network), can help “adaptive cruise control” and “forwards/rear collision warning system”, which are obviously important for the realization of fully self-driving cars. A CNN consists of multiple layers, where each layer performs multiple sets of convolutions. The convolutional filters at each layer are programmed to look for certain features through a “feature detector”. For example, horizontal lines, vertical lines, etc., are being detected by the convolution neural network.

    Implementing ML algorithms in hardware is challenging. To achieve accuracy, inference ASIC for self-driving cars needs to address the following challenges:

    Performance: A single high-definition camera that can capture 1920 × 1080 images at 60 frames per second. A car can have 10 or more of these cameras.

    Power consumption: AI inference can be a power-intensive operation, especially with large accesses to remote memory.

    Functional Safety: Functional safety issues that can spread due to various failures in the hardware must be detected.


    Zugriff

    Zugriff prüfen

    Verfügbarkeit in meiner Bibliothek prüfen

    Bestellung bei Subito €


    Exportieren, teilen und zitieren



    Titel :

    Deep Learning ASIC Design


    Beteiligte:
    Ren, Jianfeng (Autor:in) / Xia, Dong (Autor:in)

    Erschienen in:

    Erscheinungsdatum :

    10.08.2023


    Format / Umfang :

    24 pages




    Medientyp :

    Aufsatz/Kapitel (Buch)


    Format :

    Elektronische Ressource


    Sprache :

    Englisch




    ASIC design for monobit receiver

    Pok, D. / Chen, C.I.H. / Montgomery, C. et al. | Tema Archiv | 1997



    Automated Radiation Hard ASIC Design Tool

    White, M. / Bartholet, B. / Baze, M. et al. | British Library Conference Proceedings | 1993



    Structured ASIC Design for Space Systems Applications

    Bansal, J. / Orlowsky, B. / Rockett, L. et al. | British Library Conference Proceedings | 2005