Research map / Architectures and efficiency / RNNs and normalization
FlashKAT: Understanding and Addressing Performance Bottlenecks in the Kolmogorov-Arnold Transformer
Matthew Raffel, Lizhong Chen
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 |
|---|---|---|---|
| restructure backward kernel to accumulate per block before atomic add (FlashKAT) | KAT group-wise rational function · backward pass | time: 1.03 s -> 7.33 ms | Improved |
| FlashKAT kernel | KAT-T · ImageNet-1K | training throughput: 87.73 -> 6317.90 | Improved |
| FlashKAT kernel | KAT-S · ImageNet-1K | training throughput: 43.28 -> 3741.91 | Improved |
| FlashKAT kernel | KAT-B · ImageNet-1K | training throughput: 21.24 -> 1801.75 | Improved |
| FlashKAT kernel | KAT-S · ImageNet-1K | top-1 accuracy: 81.2 -> 81.4 | No clear effect |
| FlashKAT kernel | KAT-B · ImageNet-1K | top-1 accuracy: 82.3 -> 82.2 | No clear effect |
| FlashKAT kernel (block-wise sum reduction instead of atomic adds) | KAT · gradient dA | mean absolute error: 8.84×10^-2 -> 8.42×10^-4 | Improved |
| FlashKAT kernel (block-wise sum reduction instead of atomic adds) | KAT · gradient dB | mean absolute error: 9.63×10^-2 -> 9.81×10^-4 | Improved |
| increase FLOPs 8x via loops in forward kernel (ablation) | KAT rational function · forward pass | time: 4.89 ms -> 4.91 ms | No clear effect |
| increase FLOPs 8x via loops in backward kernel (ablation) | KAT rational function · backward pass | time: 1.03 s -> 1.03 s | No clear effect |
| FlashKAT kernel | KAT rational function · backward pass | HBM throughput: 1.01 -> 92.05 | Improved |
| FlashKAT kernel | KAT rational function · backward pass | Stall Long Scoreboard: 981.51 -> 2.31 | Improved |
Closest papers by meaning
- Kolmogorov-Arnold Transformer (ICLR 2025 · similarity 0.90)
- PowerMLP: An Efficient Version of KAN (AAAI 2025 · similarity 0.88)
- KAN: Kolmogorov–Arnold Networks (ICLR 2025 · similarity 0.87)
- On the expressiveness and spectral bias of KANs (ICLR 2025 · similarity 0.87)
- IceFormer: Accelerated Inference with Long-Sequence Transformers on CPUs (ICLR 2024 · similarity 0.86)
- Non-asymptotic Convergence of Training Transformers for Next-token Prediction (NeurIPS 2024 · similarity 0.86)
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.
