AI Paper Atlas

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

ICLR 2026 · Architectures and efficiency › RNNs and normalization · approach: Distributed LLM inference systems

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
add cost-model-guided tiling (+Tile) to heuristic scheduling (ablation)TileLang FlashMLA · FlashMLA, H100speedup vs Triton: 1.31 ×Improved
add warp partitioning (+Partition) after tiling (ablation)TileLang FlashMLA · FlashMLA, H100speedup vs Triton: additional 4.34×Improved
add cost-model-guided memory placement (+Alloc) (ablation)TileLang FlashMLA · FlashMLA, MI300Xspeedup vs Triton: 6.56 ×Improved
add tiling (+Tile) on top of allocation placement (ablation)TileLang FlashMLA · FlashMLA, MI300Xspeedup vs Triton: another 1.75 ×Improved
use TileLang instead of Triton for FlashMLA on H100FlashMLA kernel · MLA, H100speedup: 4.06–10.59×Improved
use TileLang instead of Triton for GEMMGEMM kernel · GEMM, H100speedup: 1.08–1.43×Improved
use TileLang instead of PyTorch for GEMM on AMDGEMM kernel · GEMM, MI300Xspeedup vs PyTorch: 0.94–1.05×No clear effect
use TileLang instead of Triton for dequantized GEMM on MI300XWINT4AFP16 GEMM · WINT4AFP16 GEMM, MI300Xspeedup vs Triton: 0.96 ×No clear effect
use TileLang instead of FlashAttention-V3 for MHAMHA kernel · MHA, H100 and MI300Xspeedup vs FA3: 0.98× and 0.96×No clear effect
use TileLang instead of Triton for chunk gated delta netChunk Gated Delta Net · Chunk Gated Delta Net, H100speedup: 1.10–1.45×Improved
use TileLang instead of Helion, Gluon, Tilus for GEMMGEMM kernel · GEMM, H100speedup: 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, H100code complexity: reducing code complexity by 77%Improved

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