Responsibly Reckless Matrix Algorithms for HPC Scientific Applications

High-performance computing (HPC) achieved an astonishing three orders of magnitude performance improvement per decade for decades, thanks to hardware technology scaling resulting in exponential the rate floating point executions, though slowing most recent. Captured Top500 list, this evolution cascaded through software stack, triggering changes at all levels, including redesign numerical linear algebra libraries. HPC simulations on massively parallel systems are often driven by matrix computations, whose execution depends their precision. Referred Jack Dongarra, 2021 ACM A.M. Turing Award Laureate, as “responsibly reckless” algorithms, we highlight implications mixed-precision (MP) computations applications. Introduced 75 years ago, long before advent architectures, MP methods turn out be paramount increasing throughput traditional and artificial intelligence (AI) workloads beyond riding wave alone. Reducing precision comes price trading away some accuracy (reckless behavior) but noncritical segments workflow (responsible so that requirements application can still satisfied. They offer a valuable performance/accuracy knob and, just they AI, now indispensable pursuit knowledge discovery simulations. In particular, illustrate impact representative applications related seismic imaging, climate/environment geospatial predictions, computational astronomy.