KernelAgent: Hardware-Guided GPU Kernel Optimization Via Multi-Agen...
Talkat KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026 · 2026-09-07..09-09·
summary
KernelAgent adds a hardware-guided optimization layer on top of an agentic multi-agent loop that runs real Triton kernels in a GPU sandbox, calls vendor profilers, and enforces a plan-first workflow with separate planning, editing, and verification steps. It claims 100% correctness across 250 L1/L2/L3 KernelBench tasks, produces 2.02x speedup on 100 L1 tasks versus earlier generated kernels and an average 1.56x versus default torch.compile, beating 65/100 tasks and reaching 89% roofline efficiency on H100; code and artifacts are open source.
AI-written summary · about
chapters (8)
- 0:00KernelAgent project
- 1:04Purpose and goals
- 2:49Harness architecture
- 8:06Kernel optimization
- 11:58Production results
- 14:28Optimization trajectory
- 15:32Cross-hardware approach
- 16:58Key learnings
related talks
19mAI Kernel Generation: What's working, what's not, what's next
13mKernelBot: The World's First Competitive GPU Programming Platform
25mLeveraging AI Ecosystem for Portable and Sustainable GPU Kernels in HPC
08mGPU Programming with Triton Kernels
32mPanel Discussion: The AI Kernel...
22mBridging the Hardware Gap With Code Harnesses on the Hugging Face Kernels Hub
18mYour Coding Agent Should Do AI System Engineering
06mThe Building Blocks of Agentic Al