By 2018, the AI software landscape faced severe fragmentation, with numerous frameworks like TensorFlow, PyTorch, and ONNX each developing distinct “AI graphs” and operational paradigms. This silos approach led to inefficiencies as different teams attempted to optimize for various hardware, resulting in duplication of efforts. Recognizing the need for a unified compiler infrastructure, a team at Google—including Chris and manager Jeff Dean—set out to create MLIR (Multi-Level Intermediate Representation). MLIR allows developers to define custom representations tailored for diverse domains, promoting modularity and composability across various AI stacks. Despite its technical success, MLIR struggled with governance and competition among companies, leading to identity confusion between being a general-purpose compiler and an AI solution. The project became a battleground for competing visions, stalling the dream of democratized AI computing. While MLIR powers a range of projects today, challenges in robust performance and integration remain prevalent in the fragmented ecosystem.
Source link
