Seminar by Dr. Sohan Lal – Harnessing the Power of GPUs: A Path to Efficiency and Sustainability

Abstract: Graphics Processing Units (GPUs) have rapidly evolved from fixed-function graphics engines to general-purpose parallel computing platforms, powering a broad spectrum of applications, from AI and scientific computing to embedded systems. Their high performance-per-watt has made them central to modern computing, including their adoption in 8 of the world’s top 10 supercomputers. Yet, GPUs continue to face challenges related to high power consumption and reduced energy efficiency under low utilization, critical hurdles on the path to sustainable and exascale computing.

In this talk, I will explore key architectural bottlenecks that limit GPU efficiency and briefly present architectural techniques (targeting the memory hierarchy) we have developed to address them. I will also highlight simulation and benchmarking tools from our group for modeling power and emerging features in NVIDIA Hopper and Blackwell architectures. The talk will conclude with an overview of our ongoing projects, such as LEAP (Locality-Driven High-Performance, Energy-Efficient GPUs) and ScaleHPC (Scalability Prediction for HPC Systems), both aimed at advancing sustainable GPU computing.

Short Bio: Dr. Sohan Lal is a computer architecture researcher with a particular focus on GPU and heterogeneous architectures. He has led the Massively Parallel Systems (MPS) research group as a Junior Professor at TU Hamburg for four years and previously worked as a postdoctoral researcher at TU Berlin, where he also earned his Ph.D. in 2019.

Dr. Lal's research has been published in top-tier venues such as IPDPS, Euro-Par, DATE, HPEC, ICCD, and IJPP, including two papers that received Best Paper and Best Poster Awards. He has successfully secured competitive third-party funding from DFG, BMBF, and ECIU.