PRIME

Table of contents

Description
Dataset
AI Principles
Acknowledgements
How to cite
Disclaimer

Description

An introductory tutorial for the PRIME algorithm is available as a Colaboratory notebook:

Dataset

We provide the PRIME dataset for nine applications, collected using an industry grade simulator. The dataset is available on Google Cloud Storage:

• PRIME Dataset: gs://gresearch/prime

You may download the dataset either by using the Google Cloud Storage web interface or using gsutil:

gsutil cp -r gs://gresearch/prime /tmp/prime/

This dataset contains both infeasible and feasible data points as described in PRIME. The descriptors of the collected data are presented in the table below.

	# of Infeasible	# of Feasible	Max Runtime (ms)	Min Runtime (ms)	Average Runtime (ms)
MobileNetEdgeTPU	384355	115711	16352.26	252.22	529.13
MobilenetV2	744718	255414	7398.13	191.35	375.05
MobilenetV3	797460	202672	7001.46	405.19	993.75
M4	791984	208148	35881.35	335.59	794.33
M5	698618	301514	35363.55	202.55	440.52
M6	756468	243664	4236.90	127.79	301.74
UNet	449578	51128	124987.51	610.96	3681.75
T-RNN Dec	405607	94459	4447.74	128.05	662.44
T-RNN Enc	410933	88880	5112.82	127.97	731.20

A demo on how to parse the dataset on Google Cloud Storage and reproducing the numbers in the table above is available as a Colaboratory notebook:

Principles

This project adheres to Google's AI principles. By participating, using or contributing to this project you are expected to adhere to these principles.

Acknowledgements

For their invaluable feedback and suggestions, we extend our gratitude to:

• Learn to Design Accelerators Team at Google Research
• Google EdgeTPU
• Vizier Team at Google Research
• Christof Angermueller
• Sheng-Chun Kao
• Samira Khan
• Xinyang Geng

How to cite

If you use this dataset, please cite:

@inproceedings{prime:iclr:2022,
  title={Data-Driven Offline Optimization For Architecting Hardware Accelerators},
  author={Kumar, Aviral and Yazdanbakhsh, Amir and Hashemi, Milad and Swersky, Kevin and Levine, Sergey},
  booktitle={International conference on learning representations},
  year={2022},
}

Disclaimer

This is not an official Google product.