Chiplets for generative AI

 

By Jawad Nasrullah, CEO - Palo Alto Electron

Generative AI models, known for their large size and substantial computational demands, are pushing the boundaries of traditional computing infrastructure. As the industry seeks solutions to mitigate costs, execution times, and the environmental impact of these models, the concept of scale-out computing traditionally seen at the data center level is being integrated into IC (Integrated Circuit) packaging using chiplet technology. This integration aims to address the challenges of power consumption and thermal design. The talk explores innovative strategies to enhance chip efficiency and reduce overheads. Key approaches include the development of AI-specific core chiplets, the implementation of efficient communication fabrics, the expansion of on-chip memory, the incorporation of more components within the IC package, the improvement of die-to-die interfaces, and the adoption of vertical chip stacking technologies. These techniques are vital for reducing power and mitigating hotspots.