Recent advances in Language Models (LMs) enable conversational systems that interpret natural-language requests, select application actions, and generate user-facing responses. However, existing studies typically either examine complete assistants with limited deployment evidence or benchmark individual components in isolation. This work advances the state of the art through a unified cross-platform characterization of a local speech-to-text-LM-text-to-speech pipeline combining Moonshine, Piper, and four quantized LMs ranging from 1.2B to 8B parameters. The pipeline is deployed on an Nvidia Jetson AGX Orin, a Raspberry Pi 5, and an STM32MP257F-EV1, and evaluated in thermostat and washing-machine use cases. Action-selection capability is assessed on 200 labelled explicit and implicit requests per use case, while deployment is characterized through metrics such as memory footprint, speech Real-Time Factor, decoding throughput, power consumption, etc. Granite 3.3 (8B) achieves the highest Macro F1-scores, whereas Qwen3.5 (4B) provides the best capability–resource trade-off. Speech processing is faster than real time on the Jetson and Raspberry Pi, while the STM32MP2 supports smaller LMs in memory but is unsuitable for responsive CPU-only conversation. LM decoding accounts for over 90% of turn duration, and memory-Roofline analysis identifies model footprint, runtime support, response length, and streaming execution as the main optimization levers.
Memory bandwidth and intermediate buffering are major bottlenecks in high-throughput FPGA-based vision systems deployed on resource-constrained edge platforms. Stereo vision reconstructs 3D information by analyzing disparities between two-dimensional images. It is a representative workload that requires evaluating many disparity hypotheses while maintaining real-time performance and low power consumption. This paper presents a memory-efficient FPGA architecture for block-based stereo matching, demonstrated using a PatchMatch formulation. The proposed design employs a compact graph-based unary cost model that enables low-bit (2-bit) cost representation, together with a single-buffer cumulative cost architecture that eliminates multi-row buffering during block aggregation. These architectural choices result in a fully streaming dataflow with significantly reduced BRAM and LUT requirements. The system is implemented on a Xilinx Zynq UltraScale+ (xczu15eg) FPGA and achieves real-time processing at 192 FPS for Full HD (1920×1080) resolution with 128 disparity levels, operating at 400 MHz. The design delivers a throughput of 51,199 MDE/s while utilizing 43.7k LUTs, 50.1k FFs, and 169.5 BRAMs, corresponding to approximately 35% reduction in LUTs and 50% reduction in BRAMs compared with the baseline Block PatchMatch architecture. These results demonstrate that the proposed architecture effectively balances accuracy and hardware efficiency, providing a scalable and resource-efficient solution for high-throughput stereo matching on FPGA-based edge systems.
Recent advances in large language models (LLMs) have demonstrated strong capabilities in program processing tasks, motivating their application to visual programming languages (VPLs). However, VPLs are inherently graph-structured and rely on syntactic, structural, and semantic constraints, whereas LLM inputs may encode these interdependent properties only partially or implicitly, creating a representation mismatch. This paper systematically reviews a corpus of 41 studies published between 2022 and 2025, focusing on LLMs for executable VPLs. We introduce a graph-and-constraint framework that captures graph structure and constraints and use it to analyze target languages, representations, methods, and evaluation practices. Our analysis shows that linearized representations remain dominant. Only four studies use Multimodal Large Language Models (MLLMs) to process images or videos, and none conducts a controlled comparison with structured representations. Evaluation remains fragmented and lacks standardization. We identify key challenges in preserving explicit relations in serialized inputs, recovering implicit relations from visual cues, integrating constraints, addressing limited data. Together, these findings clarify the limitations of current approaches and inform the development and evaluation of more reliable LLM-based methods across heterogeneous VPLs.
