Running summarizing tools on a new article
Check the paper below, then start its synthesis.
FPGA-based hardware-accelerated sentiment analysis using a hybrid LLM model with OpenVINO and intel FPGA AI suite optimization
Bonthu Chandra Praveen, K. Sivasankaran (Vellore Institute of Technology University)
Abstract: The rapid growth of social media has generated large volumes of user-generated text, making sentiment analysis important for understanding and extracting meaningful user opinions. However, efficient deployment of accurate sentiment analysis models on heterogeneous hardware platforms is challenging because of the high computational complexity. To address this, we propose a hybrid RoBERTa-BiGRU model with an attention-based fusion mechanism that combines contextual and sequential features. The proposed model achieves 96.50% accuracy and a 96.55% F1-score on the IMDb dataset, showing higher performance than existing hybrid deep learning models by approximately 3 − 4 % . For efficient, and high-performance real-time deployment, the model is converted to Open Neural Network Exchange (ONNX) and optimized using OpenVINO in FP 32 and FP 16 precision. The optimized model shows up to 4x lower latency and approximately 30x higher throughput compared to the baseline PyTorch implementation at batch size 1. Furthermore, deployment using heterogeneous CPU-FPGA execution with the Intel FPGA AI Suite demonstrates that the Agilex 5 delivers a throughput of 181.26 fps, whereas the throughput of Arria 10 is 150.72 fps. This work highlights high-performance, accurate, and efficient real-time sentiment analysis on a heterogeneous CPU-FPGA platform using hybrid modeling and hardware-aware optimization.