Panel Discussion: The AI Kernel...
summary
Kernel-level optimization uses LLMs plus tooling like Triton, Thunderkittens and Mojo to generate CUDA/HIP/PTX kernels but many auto‑generated kernels silently fail verification (examples include hardcoded softmax=1 that passes broken tests). Reports show common kernel sizes around ~300 lines per kernel, verification often requires full model validation rather than unit tests, and continuous test suites plus profilers and evolutionary test‑time scaling are needed. Labs like DeepSeek and Moonshot AI demonstrate big speedups but expect repeated human‑in‑the‑loop sanity checks; broad automation looks plausible within a decade not a year.
AI-written summary · about
chapters (11)
- 0:00AI kernel revolution panel
- 1:24Panelists and backgrounds
- 3:09Why RL hasn't produced great kernels
- 8:01Verification challenges in kernels
- 12:16Compilers hindering kernel development
- 16:18Superhuman kernel LLMs soon?
- 19:50Future of human-AI kernel writing
- 23:28Compute for kernel generation
- 27:08LLMs vs AlphaTensor for optimization
- 28:35Abstraction levels for LLM kernels
- 32:03Closing remarks
related talks
18mKernelAgent: Hardware-Guided GPU Kernel Optimization Via Multi-Agen...
18mYour Coding Agent Should Do AI System Engineering
32mKeynote: AI & the stuff built for AI - are they actually useful for data science?
50mAI IN SRE: HYPE VS REALITY IN 2026
28mAutomatic Operator Generation for Open AI frameworks: MindSpore/AKG
08mGPU Programming with Triton Kernels
40mPannel Discussion | From Hype To Value: Real World Applications Of Generative AI And LLMs
11mSimAI: Unifying Architecture Design and Performance Tuning for Large-Scale Large...