Research map / Architectures and efficiency / RNNs and normalization
TileLang: Bridge Programmability and Performance in Modern Neural Kernels
Lei Wang, Yu Cheng, Yining Shi, Zhiwen Mo, Zhengju Tang, Wenhao Xie, Tong Wu, Lingxiao Ma, Yuqing Xia, Jilong Xue, Fan Yang, Zhi Yang
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 |
|---|---|---|---|
| add cost-model-guided tiling (+Tile) to heuristic scheduling (ablation) | TileLang FlashMLA · FlashMLA, H100 | speedup vs Triton: 1.31 × | Improved |
| add warp partitioning (+Partition) after tiling (ablation) | TileLang FlashMLA · FlashMLA, H100 | speedup vs Triton: additional 4.34× | Improved |
| add cost-model-guided memory placement (+Alloc) (ablation) | TileLang FlashMLA · FlashMLA, MI300X | speedup vs Triton: 6.56 × | Improved |
| add tiling (+Tile) on top of allocation placement (ablation) | TileLang FlashMLA · FlashMLA, MI300X | speedup vs Triton: another 1.75 × | Improved |
| use TileLang instead of Triton for FlashMLA on H100 | FlashMLA kernel · MLA, H100 | speedup: 4.06–10.59× | Improved |
| use TileLang instead of Triton for GEMM | GEMM kernel · GEMM, H100 | speedup: 1.08–1.43× | Improved |
| use TileLang instead of PyTorch for GEMM on AMD | GEMM kernel · GEMM, MI300X | speedup vs PyTorch: 0.94–1.05× | No clear effect |
| use TileLang instead of Triton for dequantized GEMM on MI300X | WINT4AFP16 GEMM · WINT4AFP16 GEMM, MI300X | speedup vs Triton: 0.96 × | No clear effect |
| use TileLang instead of FlashAttention-V3 for MHA | MHA kernel · MHA, H100 and MI300X | speedup vs FA3: 0.98× and 0.96× | No clear effect |
| use TileLang instead of Triton for chunk gated delta net | Chunk Gated Delta Net · Chunk Gated Delta Net, H100 | speedup: 1.10–1.45× | Improved |
| use TileLang instead of Helion, Gluon, Tilus for GEMM | GEMM kernel · GEMM, H100 | speedup: 1.15 ∼1.62×, 1.52 ∼1.83×, 1.87 ∼2.12× | Improved |
| use TileLang instead of ThunderKittens for GEMM (code size) | GEMM kernel · GEMM, H100 | code complexity: reducing code complexity by 77% | Improved |
Closest papers by meaning
- DRTriton: Large-Scale Synthetic Data Driven Reinforcement Learning for Triton Kernel Generation (NeurIPS 2026 · similarity 0.83)
- EGG: An Expert-Guided Agent Framework for Kernel Generation (ICML 2026 · similarity 0.83)
- SGLang: Efficient Execution of Structured Language Model Programs (NeurIPS 2024 · similarity 0.83)
- FlashRNN: I/O-Aware Optimization of Traditional RNNs on modern hardware (ICLR 2025 · similarity 0.83)
- KernelBench: Can LLMs Write Efficient GPU Kernels? (ICML 2025 · similarity 0.83)
- KLay: Accelerating Arithmetic Circuits for Neurosymbolic AI (ICLR 2025 · similarity 0.83)
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.
