Google published a repository called Accelerator Agents, a pair of Gemini powered tools meant to help developers shift machine learning workloads onto Cloud TPUs. One piece, MaxCode, is built to translate PyTorch models into JAX, the framework Google’s chips run natively. The other, MaxKernel, handles the lower level work of drafting and tuning Pallas kernels, JAX’s language for writing custom accelerator code, and can convert existing CUDA kernels so they run on TPU hardware instead.
That conversion step is the real obstacle standing between developers and switching chip vendors. Most production models are built on Nvidia’s CUDA ecosystem, and reimplementing the math for a different architecture is slow, error prone work that few teams choose to do by hand. Automating it lowers the cost of leaving CUDA, which makes this release a competitive move rather than a developer convenience.
Google’s own documentation describes both tools as still evolving, and MaxCode specifically as early stage with limited coverage of model types. The project carries no independent benchmark results, and Google labels it unsupported rather than a finished product ready for production use.
Google, via its AI-Hypercomputer repository on GitHub, published this month.
