Reproduce “Gemini: Mapping and Architecture Co-exploration for Large-scale DNN Chiplet 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

Jingwei Cai, Zuotong Wu, Sen Peng, Yuchen Wei, Zhanhong Tan, Guiming Shi, Mingyu Gao, Kaisheng Ma

The Gemini framework co-optimizes architecture and mapping for large-scale Deep Neural Network (DNN) chiplet accelerators, integrating monetary cost into the design process. This approach achieves average improvements of 1.98x in performance and 1.41x in energy efficiency compared to state-of-the-art baselines, with a 14.3% increase in monetary cost.

Key insights

  • Moderately partitioning an accelerator into chiplets can reduce monetary cost while maintaining performance and energy efficiency, whereas overly fine-grained partitions degrade all metrics due to excessive D2D overheads.
  • Optimal core granularity exists where the Energy-Delay Product (EDP) initially improves with finer granularity (longer pipelines) but eventually declines due to diminishing returns and utilization loss.
  • Reusing common chiplet designs for accelerators of varying computational power is feasible with careful co-design, offering a significant reduction in Non-Recurring Engineering (NRE) costs with acceptable performance overheads.

Summary by alphaXiv.

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