Intel Announces Day 0 Support and Multi-GPU Deployment for MiniMax H3 Video Model
Intel has announced Day 0 support and a multi-GPU deployment solution for the open-source multimodal video model MiniMax H3 on Intel Arc Pro B70 GPUs. The solution leverages hybrid parallelization and offloading optimizations to accelerate the entire video generation pipeline. This deployment showcases the capability of Intel's Arc Pro GPUs to handle demanding generative AI workloads, offering a scalable alternative to dominant hardware options. It also demonstrates how advanced parallel computing techniques can overcome hardware memory limits for large-scale video models. The deployment uses 8-card Tensor Parallelism (TP) for the text encoder, 8-card Ulysses Sequence Parallel (USP) with layer-wise offloading for the Diffusion Transformer (DiT), and a hybrid 2TP×4USP scheme to balance computation and communication overhead. Additionally, key operators like matrix multiplication and attention were optimized using Intel's llm-scaler-omni project.
## BACKGROUND
MiniMax H3 is an open-weights multimodal model capable of generating 2K resolution videos with stereo audio from text, images, and audio inputs. Ulysses Sequence Parallelism (USP) is a parallel computing method that splits the sequence dimension of transformer models across multiple GPUs to handle long-context inputs efficiently.