Feeding the Beast: Power Delivery and Runtime Control for Next-Gen AI Accelerators by Dr. Jawad Haj-Yahya

Abstract

Modern AI accelerators, GPUs, and specialized silicon are no longer limited solely by transistor density or memory bandwidth; they are increasingly bottlenecked by the physics of power delivery and by how intelligently systems manage that power in real time. With current demands in high-end AI hardware already scaling past 1,500 Amperes at sub-volt levels, delivering clean, stable power has become a primary architectural constraint. This talk traces the journey of energy from the datacenter infrastructure to the chip’s final metal layers, focusing on the “last few centimeters” of the power delivery network, where the most severe efficiency, voltage, and performance challenges occur. We will examine package-level resistance, on-die IR drop, and fast di/dt transients to understand how voltage droop directly impacts system-level performance, predictability, and workload throughput. Beyond the physical PDN, the talk will explore the growing role of software and algorithmic control in power delivery. Modern AI systems increasingly rely on runtime telemetry, firmware control loops, workload-aware scheduling, DVFS policies, predictive droop mitigation, and power-aware orchestration to shape demand before electrical limits are reached. AI workloads themselves — through batching, sparsity, synchronization, memory bursts, and phase behavior — create distinctive power signatures that must be understood across the hardware/software stack. Finally, the talk highlights the critical need for close co-design between microarchitects, power engineers, system software teams, and algorithm designers. We will discuss how adaptive circuits, hardware-managed predictive mechanisms, runtime power management, and emerging technologies such as backside power delivery are reshaping the future of silicon scaling.

About the Speaker

Dr. Jawad Haj-Yahya is a Principal Architect at Meta Switzerland with over two decades of experience in computer architecture and power management. A recipient of the prestigious Intel Achievement Award, he has previously held senior research and architectural roles at Rivos Inc., Intel, Huawei, ETH Zurich, A*STAR, and NTU Singapore. Dr. Haj-Yahya specializes in energy-efficient systems and hardware security, holds over 10 US patents, and is a published author in leading journals and conferences, including ISCA, MICRO, and HPCA.