A new algorithm called FlashSinkhorn 2 can match up two enormous sets of simulated particles against each other in a few hours on one GPU, a job that normally chokes on sheer data volume.
The method tackles entropic optimal transport, the math used to find the cheapest way to match one set of points to another. Earlier GPU solvers, including the original FlashSinkhorn, still had to check every possible pairing between points in each pass, which gets expensive fast as datasets grow. FS2 instead works in two stages: a coarse pass groups nearby points into cells and solves on those cell centers first, only dropping down to individual points where a sampling check shows the coarse answer isn't good enough. A second, block-sparse pass then cleans up the remaining error. On an A100 GPU, the researchers used it to match two particle sets of 134 million points each, drawn from a cosmological N-body simulation, hitting their accuracy target in under 2.5 hours.
That scale matters because optimal transport shows up anywhere two large point sets need comparing: cosmology simulations, single-cell biology data, 3D point clouds from sensors. Most existing tools choke on datasets this size or need accuracy trade-offs the researchers avoid here. The team says this is the largest discrete problem of its kind solved to this precision in hours, not days.
The code is open-source, which is the part worth watching: a faster, cheaper tool tends to spread into other fields faster than a faster, closed one. Whether FS2 actually displaces existing solvers depends on how it holds up outside synthetic benchmarks and cosmology.