AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

FlashSinkhorn: IO-Aware Entropic Optimal Transport on GPU

Felix X.-F. Ye, Xingjie Li, An Yu, Ming-Ching Chang, LINSONG CHU, Davis Wertheimer

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Differentiable optimization layers · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
fuse Sinkhorn updates into streaming Triton kernels with online LSE (vs KeOps generic map-reduce)FlashSinkhorn · forward pass n=m=10,000, d=64runtime (ms): 125.5 -> 8.2Improved
fused streaming kernels (vs KeOps) reduce kernel launchesFlashSinkhorn · NCU profiling forward passkernel launches: 854 -> 130Improved
streaming kernels instead of tensorized dense matrixFlashSinkhorn · forward pass n=m=10,000, d=64HBM traffic (GB): 98 -> 0.08Improved
streamed backward pass with cached normalization statistics (vs KeOps)FlashSinkhorn · high dimension d>=512 forward+backwardspeedup: 100-200x speedups at d>=512Improved
use streaming HVP kernels in the CG loopFlashSinkhorn · HVP, d=64 and d=128speedup over KeOps: 3-6x at d=64, up to 25x at d=128Improved
lower epsilon from 0.10 to 0.01 (ablation)FlashSinkhorn · n=m=10000, d=64, 10 iterations10-iteration forward time: 16-17x speedup over KeOps maintainedNo clear effect
lower epsilon from 0.10 to 0.01 (ablation)FlashSinkhorn fp32 · relative error vs fp64 dense referencerelative error: 4x10^-5 -> 7.7x10^-4Hurt
use tensorized baseline at small n or very high dtensorized backend · small n (below ~2000)runtimeMixed
choose symmetric vs alternating fused update (ablation)FlashSinkhorn · small n vs large n and high druntimeMixed
switch from Adam to Newton once Lanczos detects saddle escapeOT-based shuffled regression · flow cytometry, n=40000wall time (s): 2785s -> 2423sImproved
stream OTDD cost with label lookup instead of tensorized backendOTDD MNIST to Fashion-MNIST · d=512peak memory: under 1 GB at n=60000; tensorized OOM beyond n=20000Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy