Discussion of: DOE/NSF Workshop on Correctness in Scientific Computing

David:

• Bottom of page 5: "The overall approach taken to achieve end-to-end correctness will typically involve the

use of uncertainty quantification (e.g., to account for noisy data), checks of agreement with real-world observations (e.g., pointwise validation against real-world measurements), and statistical methods (e.g., similar to those used to provide projections for hurricane trajectories)."

• Top of page 7: "The opportunity costs of scientific computing software bugs can be arbitrarily high. If a bug surfaces during an experiment that cannot be repeated (e.g., processing a rare natural phenomenon), the world community

stands to lose. This is one reason why we absolutely must have hardened and trustworthy components at our disposal. Unfortunately, practical experience (e.g., inability to handle “difficult matrices [132]”) suggests that we are really not prepared in this manner."

• Middle of 3rd paragraph on page 7: "... novel types of interconnect technologies ..."

• "last-mile neglect”

• "...loating-point details of GPUs—especially Tensor Cores—can vary in subtle ways with respect to their rounding behaviors..."

• " Ideally, “compatible” must mean well below—meaning, the numerical errors must ideally be a small fraction of the modeling error. While this fact is somewhat well-known, studies/tools to check whether such relationships hold are lacking..."

• Scientific computing: "First is the oracle problem: for many scientific applications, the correct result on inputs of interest is not known. "