On July 11,Pretty Mother in law Intel announced the launch of the Habana Gaudi 2, a deep learning accelerator card designed specifically for the Chinese market. The move aims to enhance AI training and inference capabilities, delivering improved performance and efficiency for deep learning tasks in China. The accelerator and its HL-225B mezzanine card build upon the architecture of the first-generation Gaudi. It features 24 programmable Tensor Processing Cores, 21 Gbps Ethernet interfaces, 96GB HBM2E memory capacity, and a total memory bandwidth of 2.4TB/s. The accelerator is optimized to support the inference needs of large-scale multimodal models. Digital solutions providers including Inspur, H3C, and XFUSION are expected to launch server products incorporating Gaudi 2. Gaudi 2 was previously released overseas last year, and this customized version is tailored to meet regulatory requirements in the Chinese market. Intel says that compared to other products targeting large-scale generative AI and language models, Gaudi 2 helps customers reduce expenses while achieving optimal performance. [Sina, in Chinese]
