Master thesis presentation: Efficient Mamba-Based Accelerator Architecture for Massive-MIMO Indoor Localization Using the LuViRA Dataset

High-accuracy indoor localization is becoming important for applications such as robotic navigation, emergency healthcare, and smart transportation. This thesis investigates a hardware-friendly Slim-Mamba model for Massive-MIMO indoor localization using the LuViRA dataset, where short sequences of wireless observations are used to capture temporal changes in the channel that arise from multipath propagation and variations in the indoor environment. A simplified Mamba-style model is first developed and then implemented as a Zynq MPSoC-based FPGA accelerator with processor-side control, AXI-stream dataflow, a shared reconfigurable MAC fabric for projection layers, and a FIFO-based state-space model pipeline. A fixed-point implementation flow is developed from Python to bit-true C++ and RTL. The floating-point Slim-Mamba model achieves a mean localization error of 0.135 m, while the fixed-point hardware model achieves 0.141 m with an output MAE of 0.021 compared with the floating-point output. The FPGA design closes timing at 100 MHz and reaches an estimated throughput of 471 frames/s, exceeding the 100 frames/s real-time target. On-board validation confirms that the processor control, memory movement, streaming interfaces, and FPGA computation operate together as a complete prototype system.

Supervisors: Ilayda Yaman, Sijia Cheng 

Examiner: Pietro Andreani