Intel optimizes deployment of open MiniMax H3 video model across multiple GPUs

Intel announced immediate support for the multimodal video generation model MiniMax H3, an open-source project, with optimized implementations on its Xe Flex GPU architecture. The company demonstrated a parallelization strategy to distribute processing across multiple graphics units. The developed solution partitions the model into specialized components: the encoder leverages tensor parallelism for text processing, while the main decoder employs layer-wise offloading techniques to overcome memory constraints. The final decoding stage runs on a single GPU. Beyond this initial approach, the team engineered a hybrid parallelization scheme that reduces communication overhead and improves load balancing between computation and data transfer, enhancing overall performance for large-scale video generation.