Analysis of Different Inference Implementations for Deep Learning Model on ADITYA-U Tokamak

Data driven deep learning model is implemented on ADITYA-U with precursor of disruption event using various important diagnostic signals. Model proves useful with simulated data and validated for previous shots with unseen data. Model output needs to be deployed on ADITYA-U disruption mitigation system with real-time hardware using appropriate inference execution time. In order to validate suitability and execution time, the benchmark results need to be analyzed for various methodologies. LSTM-based model inference is developed using various methods like Python, frugally deep, Keras2cpp, and kears2c. The setup is made to feed analog signal input at every 1 kHz rate, and the inference is made using various methods to benchmark results with statistics. Same platform is used to get statistics and analyzed for repetitive cycles to validate results. The keras2c inference gives optimized result in range of below 100 us time which is best fit for real-time implementation for deployment on ADITYA-U. This paper reports the analysis of benchmark result for the deep learning model implemented for ADITYA-U tokamak.