Communication-and-computation integrated designs of next-generation intelligent edge : resource management, wirelessly powered learning, and in-memory baseband processing

Abstract

The next-generation mobile communication networks are envisioned to provide high-quality services for artificial intelligent (AI) and Internet-of-Things applications at the network edge. For the implementation of edge intelligence, federated edge learning (FEEL) is a popular edge learning framework. One of the challenges faced by FEEL is the energy bottleneck, i.e., training and transmission of AI models on energy and resource constrained devices. On the other hand, the sixth generation (6G) communications require ultra-fast and energy-efficient baseband processing, while traditional complementary metal-oxide-semiconductor (CMOS)-based baseband processors cannot satisfy the requirements due to the challenges in transistor scaling and the von Neumann bottleneck. This dissertation addresses the two issues by integrated computation-and-communication (C2) designs.

First, we contribute to the energy-efficient implementation of FEEL by designing joint C2 resource management (C2RM). The design targets CPU-GPU heterogeneous computing architecture. To minimize the sum energy consumption of devices, we propose a novel C2RM framework featuring multi-dimensional control including bandwidth allocation, CPU-GPU workload partitioning and speed scaling at each device, and C2 time division for each link. The key component of the framework is a set of equilibriums in energy rates with respect to different control variables that are proved to be necessary and sufficient conditions for optimality. The results are applied to designing efficient algorithms for computing the optimal C2RM policies. Based on the equilibriums, we further design energy-efficient schemes for device scheduling and greedy spectrum sharing that scavenges spectrum holes resulting from heterogeneous C2 time divisions among devices.

Next, we propose the solution of powering devices using wireless power transfer (WPT) to tackle the energy bottleneck in FEEL. Focusing on the wirelessly powered FEEL systems, we derive the tradeoff between model convergence and the settings of power sources in two scenarios: the transmission power and density of power-beacons if they are deployed, or otherwise the transmission power of a server. The analytical framework relates the accuracy of distributed stochastic-gradient estimation to the WPT settings, the randomness in wireless links, and devices’ computation capacities. Furthermore, the local-computation at devices is optimized to efficiently use the harvested energy for gradient estimation. The resultant learning-WPT tradeoffs reveal the scaling laws of the model-convergence rate with respect to the transferred energy and the devices’ computational energy efficiencies. The results provide useful guidelines on WPT provisioning to warrant learning performance.

Last, to address the challenges in traditional CMOS-based baseband processors, we propose in-memory computing-based designs using resistive random-access memory (RRAM). We present and demonstrate RRAM-based in-memory baseband processing for the multiple-input-multiple-output orthogonal frequency-division multiplexing (MIMO-OFDM) air interface. Its key feature is to execute the key operations, including discrete Fourier transform and MIMO detection using linear minimum mean square error and zero forcing, in one-step. In addition, RRAM-based channel estimation as well as mapper/demapper modules are proposed. By prototyping and simulations, we demonstrate that the RRAM-based full-fledged communication system significantly outperforms its CMOS-based counterpart in terms of speed and energy efficiency. The results pave a potential pathway for RRAM-based in-memory computing to be implemented in the era of 6G mobile communications.