AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

FlashKAT: Understanding and Addressing Performance Bottlenecks in the Kolmogorov-Arnold Transformer

Matthew Raffel, Lizhong Chen

AAAI 2026 · Architectures and efficiency › RNNs and normalization · approach: Scalable approximation and equivariant compute · 5 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
restructure backward kernel to accumulate per block before atomic add (FlashKAT)KAT group-wise rational function · backward passtime: 1.03 s -> 7.33 msImproved
FlashKAT kernelKAT-T · ImageNet-1Ktraining throughput: 87.73 -> 6317.90Improved
FlashKAT kernelKAT-S · ImageNet-1Ktraining throughput: 43.28 -> 3741.91Improved
FlashKAT kernelKAT-B · ImageNet-1Ktraining throughput: 21.24 -> 1801.75Improved
FlashKAT kernelKAT-S · ImageNet-1Ktop-1 accuracy: 81.2 -> 81.4No clear effect
FlashKAT kernelKAT-B · ImageNet-1Ktop-1 accuracy: 82.3 -> 82.2No clear effect
FlashKAT kernel (block-wise sum reduction instead of atomic adds)KAT · gradient dAmean absolute error: 8.84×10^-2 -> 8.42×10^-4Improved
FlashKAT kernel (block-wise sum reduction instead of atomic adds)KAT · gradient dBmean absolute error: 9.63×10^-2 -> 9.81×10^-4Improved
increase FLOPs 8x via loops in forward kernel (ablation)KAT rational function · forward passtime: 4.89 ms -> 4.91 msNo clear effect
increase FLOPs 8x via loops in backward kernel (ablation)KAT rational function · backward passtime: 1.03 s -> 1.03 sNo clear effect
FlashKAT kernelKAT rational function · backward passHBM throughput: 1.01 -> 92.05Improved
FlashKAT kernelKAT rational function · backward passStall Long Scoreboard: 981.51 -> 2.31Improved

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