Research map / Large language models / Agents, code and math
KernelFoundry: Hardware-Aware Evolutionary GPU Kernel Optimization
Nina Wiedemann, Quentin Leboutet, Michael Paulitsch, Diana Wofk, Benjamin Ummenhofer
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 gradient guidance to the evolutionary search (ablation) | KernelFoundry (SYCL, 20 iters) · KernelBench L2 repr. set (SYCL) | average speedup: 2.178 -> 2.539 | Improved |
| add meta-prompt evolution (ablation) | KernelFoundry (SYCL, 20 iters) · KernelBench L2 repr. set (SYCL) | geometric speedup: 1.600 -> 1.833 | Improved |
| add meta-prompt evolution (ablation) | KernelFoundry (SYCL, 20 iters) · KernelBench L2 repr. set (SYCL) | average speedup: 2.505 -> 2.539 | No clear effect |
| add template-based parameter optimization (2 iterations) (ablation) | KernelFoundry (CUDA, o3-mini) · KernelBench L1 repr. set | average speedup: 1.204 -> 1.241 | Improved |
| add template-based parameter optimization (ablation) | KernelFoundry (CUDA, o3-mini) · KernelBench L2 repr. set | average speedup: 2.051 -> 2.104 | Improved |
| add template-based parameter optimization (ablation) | KernelFoundry (CUDA, GPT-o3/o4-mini/4.1) · Robust-kbench (12 tasks) | geometric mean speedup: 5.295 -> 5.225 | No clear effect |
| use KernelFoundry instead of OpenEvolve at equal budget | SYCL kernel generation, 10 iters · KernelBench L2 repr. set (SYCL) | average speedup: 1.483 -> 2.059 | Improved |
| use KernelFoundry instead of OpenEvolve with 40 iterations | SYCL kernel generation · KernelBench L2 repr. set (SYCL) | average speedup: 2.535 -> 2.732 | Improved |
| optimize kernels on the target GPU rather than another GPU | KernelFoundry kernels · KernelBench L2 repr. set, LNL target | average hardware-speedup: 1.537 (LNL-optimized on LNL) | Improved |
| KernelFoundry instead of best published robust-kbench kernels | CUDA kernels · Robust-kbench (12 tasks) | geometric mean speedup: 3.499 -> 5.295 | Improved |
| generated SYCL kernels vs oneDNN on matmul with relu post-op | KernelFoundry · oneDNN comparison | speedup: 0.35 | Hurt |
Closest papers by meaning
- CUDABench: Benchmarking LLMs for Text-to-CUDA Generation (NeurIPS 2026 · similarity 0.87)
- StitchCUDA: An Automated Multi-Agents End-to-End GPU Programing Framework with Rubric-based Agentic Reinforcement Learning (ICML 2026 · similarity 0.87)
- EGG: An Expert-Guided Agent Framework for Kernel Generation (ICML 2026 · similarity 0.85)
- Beyond the Trial-and-Error Loop: Hybrid Projection and Automated Tuning for Distributed Training (NeurIPS 2026 · similarity 0.85)
- Demystifying Cost-Efficiency in LLM Serving over Heterogeneous GPUs (ICML 2025 · similarity 0.85)
- KernelBench: Can LLMs Write Efficient GPU Kernels? (ICML 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.
