MaxKernel: Agentic Kernel Generation for TPUs

MaxKernel is a multi-agent system for generating and optimizing TPU kernels in JAX and Pallas, using compiler feedback and XProf hardware traces rather than relying on static, zero-shot code generation. Its specialized agents plan algorithms, write and repair kernels, synthesize tests, autotune tile configurations, verify numerical equivalence, and profile bottlenecks. The framework supports three workflows: Human-in-the-Loop collaboration, a closed-loop Autonomous agent, and Graph-Based Autonomous Search using parallel or beam exploration to avoid local optima. Powered in these experiments by Gemini 3.1 Pro and evaluated on JAXBench’s 50 diverse TPU kernel tasks, MaxKernel’s parallel search achieved a 1.58× geometric-mean speedup over the XLA baseline, with 50/50 compilation and correctness. On eight production kernels with human-written references, it reached a 2.32× geometric-mean speedup versus 2.02× for hand-tuned implementations. The gains extend to real workloads from models including Qwen3-Next and DeepSeek-V4: generated kernels accelerated the Qwen3-Next Gated DeltaNet training step by 4.70× and DeepSeek-V4 Sparse Attention prefill by 7.85× over JAX. MaxKernel also reduced Multi-Head Latent Attention latency by 8.68 percent and resolved Ragged Paged Attention crashes by inserting protective ALU clamps, showing that agentic search can address both performance and reliability in low-level accelerator programming.