Intel Arc Pro B70 First to Support Open-Source Video Model MiniMax H3

Intel announced Day 0 support for the new open-source multimodal video model MiniMax H3, with its Arc Pro B70 GPU being the first to adapt. The team implemented a multi-GPU deployment solution using 8-card tensor parallel for the encoder, 8-card USP with layer-wise offloading for DiT, and VAE on B70 GPU, accelerating the entire video generation pipeline.

DiT, the most compute-intensive module, uses 8-card USP with layer-wise offloading to dynamically load model parameters, breaking single-GPU memory limits and supporting larger models. They also developed a 2TP×4USP hybrid parallel scheme, which reduces communication overhead compared to pure 8-card USP, balancing compute and communication for better inference performance. Additionally, they optimized key operators like matrix multiplication and attention using the llm-scaler-omni project's XPU kernels, further reducing end-to-end latency.