Power Delivery for Datacenter AI Applications from Grid to Chip

#application #data-center #energy #energy-resources #data-centers #electronics-industry

Moore's Law, which is based on planar 2D models for silicon devices, postulates that the number of transistors on a microchip doubles approximately every two years. This principle has driven the advancement of smaller, faster computing devices with enhanced performance over the past sixty years. The introduction of both 2.5D and 3D packaging to support artificial intelligence has facilitated "More than Moore," increasing computing performance at faster rates. This development leads to Huang's Law, which asserts that the performance of GPUs will double every two years. As a result, the progression of computing performance to support artificial intelligence continues by transitioning from a 2D model to 2.5 and 3D models. 

As artificial intelligence (AI) workloads become larger and more complex, processing components designed to handle this data require higher levels of power. Delivering this power efficiently and reliably, while maintaining signal integrity and avoiding thermal issues, presents significant design and manufacturing challenges in both the semiconductor industry and the power electronics industry.

Meeting market demands for higher performance in smaller spaces involves addressing the power delivery issue not only in terms of efficiency but also in packaging, materials, components, and system integration. Higher power consumption impacts the power delivery system at multiple levels, including the chip level, rack level, and grid level. After power is generated, it is delivered to the data center, where it is redistributed to
racks, then to xPUs, and finally to the individual sections of the xPUs. 

Due to the unprecedented amounts of power that must be delivered, the voltage and current ratings of the various sections along the power delivery network need to be redefined in terms of higher voltage and lower current for sections that are not immediately local to the xPUs, and higher currents and lower voltage in the final millimeters or micrometers of the power delivery system. With these higher power delivery requirements, system response times to load changes become critical in all sectors of the power delivery system. The success of 2.5 and 3D packaging, as well as heterogeneous integration, drives component design to prioritize system integration rather than solely relying on standalone component parameters. 

This presentation will cover emerging trends in packaging and component design methodologies, including "signal is lateral and power is vertical," liquid cooling, distributed energy resources, modular power architecture, and sustainability. These advancements are essential for the power delivery network to meet the increasing power demands of data centers supporting AI.