Tuesday, May 26, from 5:00pm to 6:10pm
For decades, high-performance computing has operated on an implicit rule: every few years, replace the machine to get an even larger machine. But with the end of Moore and Koomey's law, is it still economically viable? In an era where machine production gets delayed, RAM prices exploding and growing scrutiny over environmental impact including procurement, is this upgrade reflex still justifiable? If yes for how long? Given the geopolitical context, sovereignty discussions and the spread of origin for machine components, do we have a choice but to extend the life of supercomputers?
In this panel we discuss past experience on increasing longevity of supercomputers, whether it was unplanned (Summit@ORNL when Aurora was late, or Sierra@LLNL confirming that El Capitan's technology was ready for production), or strategically planned due to concerns about resource usage (Fugaku@Riken). We will present challenges and how these can open new research directions and new opportunities.
We will examine the uncomfortable realities: vendor lock-in, software fragility, rising failure rates, and the hidden costs of keeping aging systems alive. But these realities also include the question of keeping a constant volume of compute, contrary to the trend of increasing applications by orders of magnitude every 6years.
Panelists
- Al Geist, Oak Ridge National Laboratory.
Bio: Al was the Chief Technology Officer and a part of the Senior Leadership Team for the DOE Exascale Computing Project. He was also the Chief Technology Officer of the Oak Ridge Leadership Computing Facility for 20 years until 2023 and was the author of the Project Execution Plan and Acquisition Plan documents for Jaguar, Titan, Summit, and Frontier system acquisitions.
- Judith Hill, Lawrence Livermore National Laboratory.
Bio: Judith is the Deputy Principal Associate Director for High Performance Computing in the Computing directorate and the Associate Program Director for Livermore Computing in the Weapon Simulation and Computing program at LLNL. She leads all aspects of Livermore Computing’s (LC) mission to support HPC environments for the LLNL and NNSA mission, including classified and unclassified operations, services, procurements, and long-term strategy.
- Fumiyoshi Shoji, RIKEN.
Bio: Fumiyoshi is currently a division director at the operations and computer technologies division, RIKEN Center for Computational Science, which is responsible for the operation and enhancement of the Fugaku and the facilities, including substations, chillers, gas turbine power generators, air handlers etc.
|
