Intel releases Habana Gaudi2 for the Chinese market

This year, Intel introduced its second generation of deep-learning chips, Habana Gaudi2 and Habana Greco, designed respectively for AI training and AI inference. These specialized processors, tailored for deep-learning applications, are built upon the efficient architecture of Habana Labs, delivering superior AI training and inference performance for clients. They will be deployed in data center applications such as computer vision and natural language processing. Both chips were birthed from Habana Labs, acquired by Intel in 2019 for two billion dollars.

Recently, at the AI Product Strategy and Gaudi2 Launch Event, Intel announced the official release of Habana Gaudi2 for the Chinese market, targeting the realms of artificial intelligence and high-performance computing. It is anticipated that manufacturers such as Inspur, New H3C, and xFusion will introduce AI servers equipped with Gaudi2. The company asserts that the Gaudi2 offers superior training performance than Nvidia’s A100 accelerator, and with a more affordable price, it can serve as a viable alternative.

Compared to its predecessor, Gaudi, the new Gaudi2 flaunts a process shrink from 16nm to 7nm; its core count is enhanced from eight to twenty-four; its memory and cache are expanded from 32GB HBM2 / 24MB to 96GB HBM2e / 48MB, and bandwidth from 1TB/s to 2.45TB/s; its network connectivity transitions from ten 100GbE links to twenty-four; and TDP is increased from 350W to 600W.

According to Intel, the Gaudi2 demonstrates impressive results on AI models. It completes the training on the GPT-3 model in 311 minutes with 384 accelerators, achieving a nearly linear scalability of 95% when moving from 256 to 384 accelerators. It delivers stellar training results on computer vision models like ResNet-50 (eight accelerators) and Unet3D (eight accelerators), and on the natural language processing model BERT (eight and sixty-four accelerators). Compared to the data submitted last November, the performance of the BERT and ResNet models improved by 10% and 4% respectively, testifying to the maturity of Gaudi2 software, which continues to evolve and mature, and stays synchronized with the growing demands of generative AI and large language models.