Welcome Here  Shenzhen Mingjiada Electronics Co., Ltd.

On 7 February, DeepSeek recently released the Janus Pro model, which has attracted industry attention for its superb performance and high accuracy. The Intel® Gaudi 2D AI accelerator is now optimised for this model, which enables AI developers to achieve deployment and optimisation of complex tasks at lower cost and higher efficiency, effectively meeting the demand for inference arithmetic in industry applications and providing strong support for the landing and scaling of AI applications.

As an innovative AIGC model, DeepSeek Janus model integrates multimodal understanding and generation. For the first time, the model adopts a unified Transformer architecture, which breaks through the limitation of traditional AIGC models relying on multipath visual coding, and achieves integrated support for comprehension and generation tasks. On this basis, Janus Pro significantly improves the performance and stability of functions such as text-generated maps by optimising the training strategy and expanding the scale of the model, and provides two versions of 1 billion parameters and 7 billion parameters, demonstrating strong competitiveness in terms of performance and accuracy.

Under Janus Pro's unified Transformer architecture, the throughput level of the Vincennes model has shifted from relying on the high computing power of accelerators to making full use of the high bandwidth and large capacity of HBM memory. With a high bandwidth of 2.45TB/sec and 96GB of HBM high-capacity memory, Intel® Gaudi 2D provides powerful computational support for the Janus Pro model, enabling a significant increase in throughput performance as the batch size (Batch Size) increases in raw mapping batch processing tasks, dramatically reducing the task processing time.

At the same time, combined with the optimisation of the Intel Optimum-Habana framework, Intel® Gaudi 2D significantly improves the throughput performance and reasoning efficiency of the text-generated image task, generating up to 16 high-quality images in about 10 seconds, with excellent performance. Moreover, developers only need to adjust a few lines of code to achieve this result, which greatly reduces the development threshold and migration cost and facilitates the efficient deployment of AI applications.

 In addition, DeepSeek's distillation model is also supported on the Xeon and Gaudi platforms, which can create a better total cost of ownership (TCO) for customers and promote the universality of AI.

All along, Intel has always adhered to an open attitude and promoted the landing of AI applications through a combination of hardware and software. Currently, Intel's rich AI product portfolio, including Intel® Xeon® processors, Intel® Core™ Ultra processors, and Intel Razzle™ graphics cards, has provided support for more than 500 open source AI models in the industry, broadly covering DeepSeek, Llama 3.1, and Tongyi Thousand Questions. At the same time, Intel is deeply committed to the open ecosystem, not only through the open ecosystem software to achieve targeted optimisation, but also through the creation of the Open Platform for Enterprise AI (OPEA) to provide support for these models.