Deprecation Warning

This example is deprecated. See CiMLoop for the latest processing-in-memory examples.

Accelergy-Timeloop Processing-in-memory (PIM) Example

This repo contains a PIM example architecture and the associated constraints designed for the accelergy-timeloop DNN accelerator evaluation framework.

Citation

@INPROCEEDINGS{ISPASS20_Wu,  
author={Wu, Yannan Nellie and Sze, Vivienne and Emer, Joel S.},  
booktitle={2020 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS)},   
title={An Architecture-Level Energy and Area Estimator for Processing-In-Memory Accelerator Designs},   
year={2020},  volume={},  number={},  pages={116-118}, 
doi={10.1109/ISPASS48437.2020.00024}}

System Requirement

The following tools are needed to perform the simulations,

• Option 1: run inside a docker with the tools installed docker

• Option 2: manually install the following tools:

  • accelergy
  • timeloop
  • accelergy-table-based-plug-in
  • accelergy-cacti-plug-in

• Please add the provided PIM estimation tables using the command

  accelergyTables -r <path to this repo>/PIM_estimation_tables
  

File Structure

• arch folder: the YAML files describing the architecture and the its associated components.
• constraint folder: describes a weight stationary dataflow with no weight reload.
• PIM_estimation_tables folder: contains necessary technology data for estimating the specific example architecture
• mapper: contains the mapspace exploration knobs
• example_outputs: contains the simulation outputs of the example workload running on the architecture

Run example simulation

To run an example simulation:

timeloop-mapper arch/system_PIM.yaml arch/components/*.yaml mapper/mapper.yaml constraints/*.yaml example_layer.yaml