This talk describes the high-level architecture of the Groq TSP chip which is a special purpose accelerator for linear algebra options, with a focus on matrix multiplication which makes it suitable for applications in the domains of machine learning and HPC. A distinctive aspect of the architecture is its deterministic mode of operation and a programming model that exploits the ability to statically schedule a computation at compile time. This talk gives an overview of the TSP architecture motivated by examples written in the Haskell Haste DSL.