MAESTRO is an open-source tool for modeling and evaluating the performance and energy-efficiency of different dataflows. MAESTRO is actively developed by the Synergy Lab at Georgia Institute of Technology. For more details about MAESTRO, please visit the following links.

• MAESTRO Website
• MAESTRO Docs

We updated the hardware description file, added off-chip bandwidth added as constraint.

We added a validation folder with data for Eyeriss and MAERI from MICRO 2019 paper.

We added a direct support for GEMM layers. For more information, please take a look at here.

We updated the naming convention of mappings and the directory structure of data folder.

Latest codebase released along with MAESTRO MICRO 2019 paper.

• Felix (Sheng-Chun) Kao ([email protected])
• Geonhwa Jeong ([email protected])
• Tushar Krishna ([email protected])

• Hyoukjun Kwon (Georgia Tech, now at Facebook Reality Labs): Main developer (core framework and functionalities)
• Prasanth Chatarasi (Georgia Tech, now at IBM Research): APIs + interface to mapping optimizers.
• Felix (Sheng-Chun) Kao (Georgia Tech): Pytorch frontend + updates to cost-model/interface + GAMMA mapper
• Geonhwa Jeong (Georgia Tech): Keras frontend + debugging + website maintainer.
• Saurabh Malik (Georgia Tech, now at Microsoft): Jupyter Notebooks demo + website.

@inproceedings{maestro_micro2019,
  author    = {Hyoukjun Kwon and
               Prasanth Chatarasi and
               Michael Pellauer and
               Angshuman Parashar and
               Vivek Sarkar and
               Tushar Krishna},
  title     = {Understanding Reuse, Performance, and Hardware Cost of {DNN} Dataflow:
               {A} Data-Centric Approach},
  booktitle = {Proceedings of the 52nd Annual {IEEE/ACM} International Symposium
               on Microarchitecture, {MICRO}},
  pages     = {754--768},
  publisher = {{ACM}},
  year      = {2019},
}

@article{maestro_toppicks2020,
  author    = {Hyoukjun Kwon and
               Prasanth Chatarasi and
               Vivek Sarkar and
               Tushar Krishna and
               Michael Pellauer and
               Angshuman Parashar},
  title     = {{MAESTRO:} {A} Data-Centric Approach to Understand Reuse, Performance,
               and Hardware Cost of {DNN} Mappings},
  journal   = {{IEEE} Micro},
  volume    = {40},
  number    = {3},
  pages     = {20--29},
  year      = {2020},
}