AI
FlashSinkhorn 2: Block-Sparse Entropic Optimal Transport
Publication: cs.AI updates on arXiv.org
Author: Felix X. -F. Ye, Yu Chin Fabian Lim, Naigang Wang, Davis Wertheimer
Source published: 05/10/2026, 04:00:00 UTC
Briefing created: 05/10/2026, 04:10:25 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 05/10/2026, 04:03:25 UTC.
Briefings
FlashSinkhorn 2 (FS2) is a new GPU-based solver for squared-Euclidean cost in entropic optimal transport that solves large discrete problems to a prescribed marginal residual on a single GPU. It achieves this by coupling a coarse stage solving on cell centroids with a block-sparse fine stage, achieving all-particle marginal residuals below 0.01 for a 1.34 billion particle cosmological N-body simulation on an A100 in under 2.5 hours. The method uses Morton-ordered blocks for fused tensor-core execution and a threshold set by block masses to bound omitted tile contributions. This represents the largest discrete EOT problem solved to this accuracy within hours to date.
Key points
- FlashSinkhorn 2 (FS2) solves large discrete entropic optimal transport problems to a prescribed marginal residual on a single GPU using a coarse stage on cell centroids and a block-sparse fine stage.
- On a single A100 GPU, FS2 solves a discrete EOT problem between two 1.34 billion particle measures from a cosmological N-body simulation to an all-particle marginal residual below 0.01 in under 2.5 hours.
Context and limitations
The study covers synthetic benchmarks and one specific A100 hardware configuration; results may vary across different GPU architectures or problem scales.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 05/10/2026, 04:10:25 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiay47r4mgox2lblnz5ubffcvzjnne3wreoblzlw4aaskiflxb6324