Skip to content

Research1 min read

MaxKernel: Multi-Agent System for TPU Kernel Generation

MaxKernel introduces a multi-agent system for designing high-performance TPU kernels, leveraging LLMs and real-time compiler feedback. It includes human-in-the-loop, autonomous, and graph-based search paradigms.

By OpenSmartRoute editorial · written through the router by llm-onprem

From arXiv cs.AI - “MaxKernel: Agentic Kernel Generation for TPUs

MaxKernel is a system that implements three paradigms for TPU kernel development: a human-in-the-loop agent for collaborative design, an autonomous agent for automated optimization, and a graph-based search for global exploration.

All three paradigms utilize a shared pool of specialized sub-agents responsible for planning, implementation, self-debugging, testing, and hardware profiling. The system is evaluated on JaxBench, a suite of 50 diverse kernel tasks, and complex workloads from open-source models.

Results show that MaxKernel consistently produces optimized implementations that match expert hand-tuned baselines and improve performance across benchmarks. The system is open-sourced and available for use.

Source: https://arxiv.org/abs/2609.04523

Published Sep 7, 2026 · updated Sep 7, 2026 · 97 words

Keep reading

Related posts

More in Research