Nova: An End-to-End MLIR Compiler for Deep Learning

Nova is an end-to-end automated JIT compiler that seeks to provide complete control over the hardware mapping of deep learning models. It merges operations across different boundaries, enhances complex memory structures, and fine-tunes execution at the register level. By capturing eager executions and integrating both forward and backward passes into a unified value-semantic dialect, Nova facilitates extensive whole-graph optimizations. An Analytic Configurator is employed to systematically derive the best configurations. This compiler is detailed in a paper on arXiv (ID: 2608.00029), which discusses the difficulties of translating high-level mathematical operations to physical hardware. It notes that while high-level tensor frameworks provide flexible abstractions, their eager execution models lack comprehensive visibility and precise control over hardware and memory, a gap that Nova seeks to fill, especially for scaling deep learning models where performance is closely tied to efficient hardware use. The paper was recently submitted to arXiv.