Mike Ranjram, an assistant professor at Arizona State University, recently shared his insights on data center power delivery for the evolving AI ecosystem during a presentation to advanced packaging professionals at IMAPS’s last 2024 chapter event in Chandler, Arizona. As artificial intelligence (AI) continues to advance, it faces challenges due to its increasing power demands. In fact, power management is cited as the biggest issue hindering AI scalability today.
According to Ranjram, Arizona is projected to experience a 40% increase in peak load over the next five years, with data centers already accounting for 10 gigawatts (GW) in pending interconnect requests with the local energy provider. These escalating power needs highlight the urgency of addressing power delivery challenges for AI.
To meet the demands of AI, Ranjram emphasized the importance of ensuring that upstream converters can handle the power requirements while maintaining isolation and power factor correction. Upstream converters, which boost voltage flow, must be able to adapt to the increasing power levels and current demands associated with AI’s evolution.
Despite the ongoing challenges in power delivery, using upstream converters with higher power density can help alleviate stress on the grid. Ranjram suggested a short-term approach of transitioning from 12 volt (V) to 48V converters to address the immediate power needs of AI.
While upstream converters address power demands, point-of-load (POL) converters also play a crucial role in accommodating high-performance semiconductors in data centers. These converters, positioned close to the power load, face scaling issues in meeting rising power demands but offer benefits such as parallelization to increase available current and widespread market support.
Looking ahead, Ranjram highlighted vertical power delivery (VPD) as a potential solution to mitigate AI scaling concerns. VPD structures offer advantages such as reduced impedance and improved power efficiency by minimizing the distance currents need to travel.
Overall, the challenges of meeting AI’s power demands require innovative solutions and continuous advancements in power delivery technology. By addressing these challenges, the industry can pave the way for the future of AI innovation and development.
Source link
