Julia on TPU: Reproducible Examples and Benchmarks

This work documents a systematic investigation of the Julia programming language on Google Tensor Processing Unit (TPU) hardware, using Reactant.jl and the XLA compiler as the execution backend. It provides nine self-contained example scripts covering progressively advanced numerical workloads, together with five quantitative benchmarks comparing Julia/Reactant against Python/JAX across matrix multiplication, vector dot product, random number generation, batched matrix multiplication, and gradient computation workloads. A containerised execution environment is provided for Docker, Podman, and Apptainer, alongside full documentation of installation methods and known issues on Google Cloud TPU VMs. A key methodological finding is that Reactant/XLA kernel dispatch is asynchronous by default, and naive timing without explicit synchronisation underestimates wall-clock time by up to 57x at large problem sizes.