Recent years have seen a proliferation of specialized ML accelerators – proposed in both academia (e.g., Gemmini, FEATHER) and industry (e.g., Google TPU, Intel AMX) – which depart significantly from traditional CPU/GPU architectures. However, research on compiler and systems support for these accelerators remains sparse, largely due to the lack of mature open-source compiler infrastructures capable of targeting them from popular ML frameworks like PyTorch and JAX. Building such support involves considerable manual effort, slowing innovation and creating a gap between hardware and software research communities.

This tutorial introduces ACT (Accelerator Compiler Toolkit), an ecosystem that automatically generates essential software tooling like end-to-end compilers from high-level ISA specifications of AI accelerators. It enables rapid prototyping and evaluation of architecture-specific optimizations, supporting tighter hardware-software co-design by significantly reducing compiler bring-up effort.

In this tutorial, we will walk through compiling real-world ML models for various existing AI accelerators and demonstrate how the infrastructure can be extended to support new designs with minimal effort. We will cover key steps in the workflow – from ISA specification to software simulation and FPGA emulation – showing rapid iteration and end-to-end evaluation. The goal is to provide attendees with practical insights and tools to accelerate systems and compiler research in the evolving ML hardware landscape.