MaxKernel: Agentic Kernel Generation for TPUs
MaxKernel uses multiple LLM agents plus compiler and profiling feedback to generate TPU kernels.
The paper describes three modes: human-guided design, a fully autonomous optimization loop, and graph-based autonomous search. Each mode draws on sub-agents for planning, implementation, debugging, testing, and hardware profiling. The authors evaluate it on 50 JaxBench TPU kernel tasks and real workloads from open-source models, reporting performance that matches expert hand-tuned baselines. The agent is described as open source. HF Daily Papers' note
The paper describes three modes: human-guided design, a fully autonomous optimization loop, and graph-based autonomous search. Each mode draws on sub-agents for planning, implementation, debugging, testing, and hardware profiling. The authors evaluate it on 50 JaxBench TPU kernel tasks and real workloads from open-source models, reporting performance that matches expert hand-tuned baselines. The agent is described as open source. HF Daily Papers' note
score 5