Reproduce “Dato: A Task-Based Programming Model for Dataflow Accelerators” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Shihan Fang, Hongzheng Chen, Niansong Zhang, Jiajie Li, Han Meng, Adrian Liu, Zhiru Zhang

Dato is a programming model for dataflow accelerators that unifies explicit data communication and sharding through first-class types. It achieves high hardware utilization, such as up to 84% for mixed-precision GEMM on AMD NPU and 98% of peak throughput for GEMM on FPGA, while significantly improving developer productivity with up to 12x fewer lines of code compared to existing frameworks.

Key insights

  • Explicitly modeling data communication and sharding as first-class, statically-checked types enables precise control over dataflow and compile-time verification, mitigating common errors such as deadlocks and overflows.
  • A systematic virtual-to-physical mapping approach allows high-level dataflow patterns to be systematically optimized and mapped to diverse physical processing elements while respecting hardware constraints.
  • Combining a high-level programming interface with hardware-aware compiler optimizations effectively achieves both high performance and significantly improved developer productivity for complex, multi-kernel workloads.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.