From AI Agents to Faster Kernels: Ben Burtenshaw & Felix LeClair (AI Plumbers #2)

2026-03-06 · Source: HuggingFace · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Intermediate, medium

Summary

Ben from Hugging Face, an advocate for high-performance kernels, discusses the Kernels Hub, a new library designed to simplify the installation of optimized kernels like Flash Attention 3 and 4. This hub reduces installation time from hours to seconds, primarily benefiting users on single instances and older hardware by lowering costs and saving time. The discussion also covers the use of AI agents, specifically Claude, in generating and optimizing kernels. While initial attempts with Claude 4.5 were challenging, defining "skills" with best practices and example scripts significantly improved its performance, enabling it to generate a kernel for Diffusers on H100 that was 50% faster than the baseline. This "skill" approach also improved open-weight models like Kimi 2.5, making them more efficient in token consumption.

Key takeaway

For Computer Vision Engineers seeking to optimize model performance on diverse hardware, you should explore the Hugging Face Kernels Hub. This resource simplifies access to high-performance kernels, including those generated by AI agents, potentially reducing operational costs and extending the utility of older hardware. Consider contributing to the Kernels GitHub repository to help expand support for a wider range of hardware configurations.

Key insights

AI agents can generate optimized kernels, but structured "skills" and robust infrastructure are crucial for widespread adoption and benefit.

Principles

• Optimized kernels extend hardware lifespan.
• Standardized kernel declarations enhance education.

Method

Define agent "skills" using best practices, reference files, and example scripts to guide kernel generation, then benchmark performance with generated test sets.

In practice

• Use Kernels Hub for rapid kernel installation.
• Explore agent-generated kernels for older hardware.
• Contribute to Hugging Face GitHub "good first issues".

Topics

• High-Performance Kernels
• Hugging Face Kernels Hub
• AI Agents
• Hardware Optimization
• Machine Learning Infrastructure

Best for: Computer Vision Engineer, Machine Learning Engineer, AI Engineer, MLOps Engineer

Related on AIssential

• Talk: Kernels Deep Dive (Ben Burtenshaw) — Standardized kernel distribution and usage significantly reduce deep learning memory bottlenecks and improve accessibility.
• NVIDIA's AI Engineers: Agent Inference at Planetary Scale and "Speed of Light" — Nader Khalil (Brev), Kyle Kranen (Dynamo) — NVIDIA GTC will feature sessions with Kyle Kranen, an architect of NVIDIA Dynamo, and Nader Khalil, from NVIDIA Brev, discussing the company's rapid growth and strategic investments in AI.
• DeepSeek Just Solved AI's Billion Dollar Problem — DeepSeek's method doubles AI system utilization by optimizing memory access between prefill and decoding machines.
• Import AI 439: AI kernels; decentralized training; and universal representations — AI systems can automate and accelerate complex software optimization tasks, yielding significant performance gains.
• AgentKernelArena: Generalization-Aware Benchmarking of GPU Kernel Optimization Agents — AgentKernelArena benchmarks AI agents on GPU kernel optimization, revealing strong performance but critical generalization gaps on unseen input configurations.
• Speeding up GPU kernels by 38% with a multi-agent system — Multi-agent systems can autonomously optimize complex software, achieving significant performance gains in weeks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by HuggingFace.