Reproduce “Instruction Scheduling in the Saturn Vector Unit” 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

Jerry Zhao, Daniel Grubb, Miles Rusch, Tianrui Wei, Kevin Anderson, Borivoje Nikolic, Krste Asanovic

Saturn implements a compact RVV 1.0 vector unit for short data-parallel batches. It keeps whole vector instructions in shallow queues, then dynamically sequences element-groups across arithmetic, load, and store paths, while loads can run ahead. Its fully featured short-vector configuration exceeds 90% utilization across many kernels and outperforms Ara, though a hypothetical long-vector configuration leads on almost all kernels.

Key insights

  • Dynamic instruction scheduling and fine-granularity explicit chaining can maintain high datapath utilization in short-vector units without requiring complex register renaming or full out-of-order execution.
  • Decoupled access-execute microarchitectures are crucial for tolerating high memory latencies and exploiting memory-level parallelism in compact vector processors.
  • Modern scalable vector ISAs can be fully supported in area- and power-constrained environments by combining short hardware vector lengths with sophisticated, yet lightweight, scheduling mechanisms.

Summary by alphaXiv.

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