MICA is the first end-to-end compiler for mesh-based AI accelerators. Submitted to OSDI 2026 and developed in collaboration with leading AI chip manufacturers, it enables automatic model adaptation for next-generation AI hardware — turning weeks of manual wafer-scale scheduling into hours of automated compilation, with generated code outperforming expert hand-tuned implementations.
Key Features
MeshIR Intermediate Representation — A novel IR design using region-aware tensors and kernels that explicitly expresses locality and placement constraints, letting the compiler "understand" mesh architecture.
Spatio-Temporal Schedule Search — Efficiently explores the schedule space across millions of cores to automatically generate optimal scheduling plans.
End-to-End Compilation Stack — Complete compilation pipeline from high-level model descriptions to mesh accelerator executable code.
Results
- 3.6x performance improvement over SOTA compilers (validated on real hardware)
- 1.4x faster than expert hand-tuned optimizations
- Wafer-scale scheduling search reduced from weeks to hours
- Compiled applications run 100–200x faster than GPU
Collaborators
Leading AI chip manufacturers
MICA 是首个面向 Mesh 架构 AI 加速器的端到端编译器。投稿 OSDI 2026,与领先 AI 芯片厂商联合开发,实现下一代 AI 硬件上的模型自动适配——将原本需要数周的晶圆级调度搜索缩短到数小时,编译器生成的代码甚至超过人类专家手动调优的性能。
核心功能
MeshIR 中间表示 — 全新 IR 设计,通过 region-aware tensors 和 kernels 显式表达局部性与放置约束,让编译器"理解" Mesh 架构。
时空调度搜索框架 — 高效探索数百万核心上的 temporal-spatial schedule 空间,自动生成最优调度方案。
端到端编译栈 — 从高层模型描述到 Mesh 加速器可执行代码的完整编译流程。
成果
- 比 SOTA 编译器性能提升 3.6×(真实硬件验证)
- 比专家手写优化快 1.4×
- 晶圆级调度搜索从数周缩短到数小时
- 编译出的应用比 GPU 快 100–200×
合作方
领先 AI 芯片厂商
