Reproduce “To SIMD or Not to SIMD: Comparing RVV vs. CGRA SoCs for Non-Linear Functions in AI” 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

Sameer Deshmukh, Naoya Mizuki, Takumi Honda

The authors compare matched simulated FP32 systems pairing a matrix unit with either a vector processor or a coarse-grained reconfigurable array (CGRA). On tested SwiGLU, the CGRA reduces SiLU activation cycles by at most 7.12×; on attention softmax, the vector unit is up to 9.8× faster. These are cycle-count results for tested mappings, not wall-clock or production performance.

Key insights

  • Spatial pipelining gives the CGRA an advantage on regular, element-wise SiLU, where it can stream inputs through a mapped datapath with little repeated instruction issue.
  • Softmax’s reductions, multiple passes, data transfers, synchronization, and CGRA setup costs favor the RVV implementation in the tested design; arithmetic throughput alone does not determine performance.
  • The comparison is specific to these implementations and workloads: the softmax mapping has identifiable inefficiencies, and the study does not measure synthesized area, power, or energy.

Summary by alphaXiv.

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