Worked on GPU memory bandwidth modeling to enhance performance and cost estimation for H100 GPUs, focusing on both the Intel-tensorflow/tensorflow and Intel-tensorflow/xla repositories. Developed a dynamic HBM bandwidth model for dot fusion in TensorFlow, introducing a DMA-size-based effective bandwidth function and a lookup table to replace hardcoded device checks, increasing model flexibility. In XLA, integrated an HBM derate curve and refactored time calculations to use the new lookup table, improving accuracy for memory-bound scenarios. Utilized C++, CUDA, and cost modeling techniques to align cross-repo approaches, supporting future GPU architectures and expanding test coverage for bandwidth-sensitive workloads.