Do We Really Need All These Bits? Reliable and Sustainable Computing Beyond Double Precision
The traditional assumption that scientific computations should be performed uniformly in double precision is becoming increasingly difficult to sustain. Modern processors, and especially AI-oriented accelerators, support a growing range of numerical formats -- from fp64 and fp32 to fp16, bf16, fp8, and beyond -- with substantial differences in performance, memory consumption, and energy efficiency. Exploiting these capabilities, however, requires more than simply replacing one precision with another. Precision reductions must be supported either by rigorous numerical analysis or by computer arithmetic tools that explore lower-precision alternatives while applying appropriate numerical sanity checks. Precision should therefore be treated as an actively managed algorithmic resource rather than as a fixed implementation choice.
This talk will examine how low- and mixed-precision computing can improve the performance and energy efficiency of numerical simulations and data-intensive scientific applications without compromising the reliability of their results. I will discuss approaches that combine low-precision computational kernels with higher-precision correction and verification mechanisms, including mixed-precision iterative solvers, iterative refinement, and computer arithmetic tools. Examples from numerical linear algebra and large-scale scientific applications will illustrate when reduced precision provides meaningful benefits and when rounding errors, convergence behaviour, data movement, or hardware characteristics limit its effectiveness. More broadly, the talk will position precision within sustainable high-performance computing, where time-to-solution, energy-to-solution, and numerical accuracy must be optimized jointly across algorithms, software, and hardware.
Session Chair: Daniel Bauer