Integrated AI and wireless for sustainable mobile network evolutions

  • * Register (or log in) to the Neural Network to add this session to your agenda or watch the replay

  • Date
    11 March 2025
    Timeframe
    13:00 - 14:00 CET Geneva
    Duration
    60 minutes
    Share this session

    Mobile networks have advanced through five generations, bringing significant benefits to society. However, the continuous evolution of these networks places a substantial financial burden on network operators, as the costs associated with building new generations of networks keep increasing. Additionally, future mobile network infrastructures and wireless devices need to integrate both wireless baseband and artificial intelligence (AI) processing capabilities to support AI-native mobile networks and applications. In this context, it is essential to develop integrated AI and wireless processors and create open-source libraries that support a purely software-defined protocol stack, decouple software and hardware functionalities, and allow flexible allocation of computing power between AI and wireless baseband processing. In this talk, we will provide a detailed overview of a RISC-V-based architecture and its customized extensions, with a focus on vector processing, which efficiently handles data-parallel operations for AI and wireless baseband processing. Traditional vector architectures face challenges such as limited vector register sizes, reliance on power-of-two vector length multipliers, and architecture-specific vector permutation capabilities.

    To address these challenges, we propose a novel approach called Unlimited Vector Processing (UVP), which enhances the flexibility and performance of vector computations. UVP introduces a unique programming model that supports non-power-of-two register groupings and larger physical register files, enabling seamless management of vectors of varying lengths while reducing strip-mining overhead. This approach categorizes vector instructions into symmetric and asymmetric groups, utilizing custom load/store mechanisms to optimize execution. Furthermore, we propose a hardware implementation of UVP that includes advanced hazard detection logic, optimized pipelines for symmetric tasks such as fixed-point multiplication and division, and a detailed permutation engine for efficient handling of asymmetric operations.

    Additionally, we introduce an open-source library for AI computing units and mobile network protocol stacks, incorporating basic physical layer signal processing through our suggested RISC-V-based extensions. This open-source library enables the true decoupling of software and hardware in mobile networks and wireless devices, promoting a sustainable evolution of mobile networks where devices can be upgraded through software without the need to rebuild existing infrastructures, thus achieving “white-box” base stations and devices. By integrating AI with wireless signal processing within a unified set of computing units, we enable efficient and flexible distribution of computing resources and facilitate the development of AI-native wireless networks. 

  • Are you sure you want to remove this speaker?