Reproduce “Understanding Reuse, Performance, and Hardware Cost of DNN Dataflows: A Data-Centric Approach Using MAESTRO” 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

Hyoukjun Kwon, Prasanth Chatarasi, Michael Pellauer, Angshuman Parashar, Vivek Sarkar, Tushar Krishna

MAESTRO models how neural-network dataflows map operations across processing elements and time, making reuse and data movement explicit. It estimates communication, runtime, energy, and buffer needs from layer mappings and hardware configurations. In reported validation cases, runtime estimates had 3.9% average absolute error against MAERI RTL and published Eyeriss results; this does not establish accuracy for all models or hardware.

Key insights

  • No single dataflow style is universally optimal across all DNN operators and layers; performance and energy efficiency depend heavily on specific operator characteristics.
  • Optimal hardware design parameters (PEs, buffers, NoC bandwidth) vary significantly for early versus late layers within the same DNN model, indicating a need for layer-specific or adaptive hardware.
  • Peak throughput is not solely determined by the number of Processing Elements (PEs); NoC bandwidth is equally critical to prevent communication bottlenecks, and larger buffers do not always guarantee higher efficiency.

Summary by alphaXiv.

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