Intel and DeepSeek Revolutionize AI with Optimized Gaudi 2D Accelerators for Breakthrough Janus Pro Model

Intel has recently made headlines by announcing the optimization of their Intel® Gaudi 2D AI accelerators specifically for DeepSeek’s cutting-edge Janus Pro model. This development is set to lower the deployment barriers for AI developers by delivering complex tasks at a reduced cost and enhanced efficiency, addressing the industry's growing demand for improved inference computational capabilities.

The Janus Pro model, launched by DeepSeek, is distinguished by its adoption of a unified Transformer architecture. This innovative approach overcomes the traditional reliance on multiple visual encoding pathways in AIGC models, providing robust support for both comprehension and generation tasks. With enhancements in training strategy and scaling of the model, Janus Pro exhibits outstanding performance and stability, particularly in text-to-image tasks. Furthermore, the introduction of versions featuring 1 billion and 7 billion parameters has fortified its competitive edge in terms of both performance and precision.

A notable advancement in the Janus Pro model under the unified Transformer architecture is its efficient data throughput. Previously, throughput was heavily reliant on the accelerator’s computational power, but now it leverages the high bandwidth and large capacity of HBM memory. The incorporation of Intel® Gaudi 2D’s 2.45TB/s bandwidth and 96GB HBM ensures significant enhancements in task batch processing capabilities, noticeably reducing task processing durations. In relation to its text-to-image model, the optimization with Intel’s Optimum-Habana framework enables the generation of 16 high-quality images in roughly 10 seconds, significantly boosting task efficiency.

Additionally, DeepSeek's distilled models have also benefited from the support of Xeon and Gaudi platforms, which assists in minimizing the total cost of ownership (TCO) and advancing AI technology accessibility. Currently, Intel's extensive AI product lineup supports more than 500 open-source AI models, including DeepSeek, Llama 3.1, and others, all while fostering an open ecosystem to provide a solid foundation for enterprise AI applications.