Reproduce “STONNE: A Detailed Architectural Simulator for Flexible Neural Network 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
Francisco Muñoz-Martínez, José L. Abellán, Manuel E. Acacio, Tushar Krishna
A new open-source, cycle-accurate architectural simulator, STONNE, allows for detailed evaluation of flexible deep neural network accelerators. The simulator validated against MAERI, an existing flexible design, identified that MAERI's folding strategy led to an average compute unit utilization of only 25% across various DNNs, suggesting significant performance gains with an optimized approach.
Key insights
- Cycle-accurate simulation is crucial for accurately quantifying the performance impact of micro-architectural design choices in complex flexible accelerators.
- The 'folding' strategy, essential for processing operations larger than available compute units, can become a critical performance bottleneck if inefficiently implemented.
- Dependencies introduced by partial sum handling during folding can prevent pipelined execution, leading to significant stalls and underutilization of compute resources.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.