Guarding Numerics Amidst Rising Heterogeneity
Ganesh Lalitha Gopalakrishnan, Ignacio Laguna, Ang Li, Pavel Panchekha, Cindy Rubio-González, Zachary Tatlock · 2021
New heterogeneous computing platforms-especially GPUs and other accelerators-are being introduced at a brisk pace, motivated by the goals of exploiting parallelism and reducing data movement. Unfortunately, their sheer variety as well as the optimization options supported by them have been observed to alter the computed numerical results to the extent that reproducible results are no longer possible to obtain without extra effort. Our main contribution in this paper is to document the scope and magnitude of this problem which we classify under the heading of numerics. We propose a taxonomy to classify specific problems to be addressed by the community, a few immediately actionable topics as the next steps, and also forums within which to continue discussions.