Re:Math Breakthrough Speeds Supercomputer Simulations
by vortico on Saturday October 19, 2019 @08:32PM
(#21301597)
In theoretical computer science, you're somewhat right. Constant reductions aren't as interesting as complexity reductions but still worth publishing papers on of course. In mathematics, it's common to get 10^9 constant speedups just be applying a clever method.
In practical computer science/engineering like HPC, people are typically excited to get 10-20% speedups, because in a project with a million dollar compute budget, that'll pay their salary that year. A 3-4 speedup is almost unheard of for already-optimized codes like LAMMPS (although these numbers are a guess at the practical maximum speed), so it's probably newsworthy.
In practical computer science/engineering like HPC, people are typically excited to get 10-20% speedups, because in a project with a million dollar compute budget, that'll pay their salary that year. A 3-4 speedup is almost unheard of for already-optimized codes like LAMMPS (although these numbers are a guess at the practical maximum speed), so it's probably newsworthy.