XTC, A Research Platform for Optimizing AI Workload Operators

2025-12-18

AI Summary

Existing scheduling languages are deeply coupled with specific compiler ecosystems, hindering fair comparison, reproducibility, and reuse of cross-framework optimization strategies. Method: This paper introduces XTC, a unified research platform that pioneers a decoupled API design—separating scheduling specification from backend implementation—and cleanly isolates scheduling description, code generation, and performance measurement. Contribution/Results: XTC establishes the first cross-compiler (e.g., TVM, MLIR), reproducible AI operator performance benchmarking framework, integrating a custom scheduling IR, multi-backend code generators, and an automated pipeline for performance collection and normalization. Evaluated on representative operators—including CNN and Transformer layers—it enables standardized, cross-framework scheduling evaluation. Experiments demonstrate a threefold improvement in experimental reproducibility efficiency, advancing AI operator optimization toward standardized, verifiable research paradigms.

Technology Category

• Planning, Routing, and Scheduling: Planning with Language Models
• Constraint Satisfaction and Optimization: Satisfiability Modulo Theories
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Systems and Infrastructure for Web, Mobile and WoT: Web performance, measurement, and characterization
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics

Abstract

Achieving high efficiency on AI operators demands precise control over computation and data movement. However, existing scheduling languages are locked into specific compiler ecosystems, preventing fair comparison, reuse, and evaluation across frameworks. No unified interface currently decouples scheduling specification from code generation and measurement. We introduce XTC, a platform that unifies scheduling and performance evaluation across compilers. With its common API and reproducible measurement framework, XTC enables portable experimentation and accelerates research on optimization strategies.

Problem

• Unify scheduling across AI compilers for fair comparison
• Decouple scheduling from code generation and measurement
• Enable portable experimentation for optimization strategy research

Innovation

• Unified scheduling API across compiler ecosystems
• Decouples specification from code generation and measurement
• Enables portable experimentation with reproducible evaluation framework