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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| fuse Sinkhorn updates into streaming Triton kernels with online LSE (vs KeOps generic map-reduce) | FlashSinkhorn · forward pass n=m=10,000, d=64 | runtime (ms): 125.5 -> 8.2 | Improved |
| fused streaming kernels (vs KeOps) reduce kernel launches | FlashSinkhorn · NCU profiling forward pass | kernel launches: 854 -> 130 | Improved |
| streaming kernels instead of tensorized dense matrix | FlashSinkhorn · forward pass n=m=10,000, d=64 | HBM traffic (GB): 98 -> 0.08 | Improved |
| streamed backward pass with cached normalization statistics (vs KeOps) | FlashSinkhorn · high dimension d>=512 forward+backward | speedup: 100-200x speedups at d>=512 | Improved |
| use streaming HVP kernels in the CG loop | FlashSinkhorn · HVP, d=64 and d=128 | speedup over KeOps: 3-6x at d=64, up to 25x at d=128 | Improved |
| lower epsilon from 0.10 to 0.01 (ablation) | FlashSinkhorn · n=m=10000, d=64, 10 iterations | 10-iteration forward time: 16-17x speedup over KeOps maintained | No clear effect |
| lower epsilon from 0.10 to 0.01 (ablation) | FlashSinkhorn fp32 · relative error vs fp64 dense reference | relative error: 4x10^-5 -> 7.7x10^-4 | Hurt |
| use tensorized baseline at small n or very high d | tensorized backend · small n (below ~2000) | runtime | Mixed |
| choose symmetric vs alternating fused update (ablation) | FlashSinkhorn · small n vs large n and high d | runtime | Mixed |
| switch from Adam to Newton once Lanczos detects saddle escape | OT-based shuffled regression · flow cytometry, n=40000 | wall time (s): 2785s -> 2423s | Improved |
| stream OTDD cost with label lookup instead of tensorized backend | OTDD MNIST to Fashion-MNIST · d=512 | peak memory: under 1 GB at n=60000; tensorized OOM beyond n=20000 | Improved |
Closest papers by meaning
- Progressive Entropic Optimal Transport Solvers (NeurIPS 2024 · similarity 0.88)
- A Truncated Newton Method for Optimal Transport (ICLR 2025 · similarity 0.87)
- Log-Averaged Mirror Prox for Fast, Large-Scale Optimal Transport in Linear Space (NeurIPS 2026 · similarity 0.87)
- Safe and Sparse Newton Method for Entropic-Regularized Optimal Transport (NeurIPS 2024 · similarity 0.86)
- Accelerating Sinkhorn algorithm with sparse Newton iterations (ICLR 2024 · similarity 0.85)
- FlashMoE: Fast Distributed MoE in a Single Kernel (NeurIPS 2025 · similarity 0.85)
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.
