| Abstract: |
In light of recent export restrictions from the United States, Hong Kong faces significant challenges in acquiring high-end GPUs, alongside associated security risks. To address these issues, our proposal outlines the design of a versatile machine learning platform that supports heterogeneous GPUs and AI chips, including domestic alternatives. This platform will harness the unique computational strengths of various architectures to deliver optimal performance, flexibility, and cost-effectiveness. Key features include dynamic workload distribution across heterogeneous GPUs and AI chips, improved cluster utilization, rapid recovery from infrastructure failures, and unified image management across different architectures. By integrating diverse GPUs and AI chips from multiple vendors, this platform will significantly enhance the efficiency of training processes, fostering innovation and accelerating advancements in artificial intelligence research and applications in Hong Kong. Moreover, this initiative aligns with Hong Kong's strategic goals of technological self-reliance and security, ensuring that our AI infrastructure remains robust and adaptable in the face of global supply chain uncertainties. By leveraging local talent and resources, this project will also contribute to building a sustainable and resilient AI ecosystem, positioning Hong Kong as a leader in cutting-edge AI technologies and applications. |
