Accelerator Zoo

This repository contains the TeAAL specification and/or HiFiber loop nests representing the following accelerators:

• DSTC
• ExTensor
• Eyeriss
• Flexagon
• FuseMax
• Gamma
• GraphDynS
• Graphicionado
• OuterSPACE
• SIGMA
• SPADE
• Tensaurus
• TPUv1

We welcome contributions of other designs from the community.

System Requirements

Hardware

• x86-64 CPU
• 5GB disk space

Software

• Docker
• Web browser

Installation Instructions

Step 0: Clone the repository

Submodules must also be recursively cloned.

git clone --recurse-submodules git@github.com:FPSG-UIUC/accelerator-zoo.git
cd accelerator-zoo

[Recommended] Option 1: Docker

Step 1: Prepare your docker-compose.yaml

Copy the docker-compose.yaml.example file to a new docker-compose.yaml

cp docker-compose.yaml.example docker-compose.yaml

Edit the docker-compose.yaml with the appropriate NB_UID (the user you want to own the repository).

Step 2: Pull the Docker image

We provide two options for obtaining the docker image. Please choose one of the options listed below.

Option 1: Use docker-compose

docker-compose pull

Option 2: Build the image from source

cd ./setup/docker/fibertree-docker
make build
cd ../../..

Step 3: Start the container

Start the container, including the Jupyter lab.

docker-compose up -d

Step 4: Open the Jupyter lab

Navigate to http://localhost:8888 to open and view the repository.

Option 2: Python Virtual Environment

Step 1: Install Python 3.12

• On Linux
• On Mac via Homebrew

Step 2: Create the virtual environment

python3.12 -m venv env

Start the Virtual Environment

source env/bin/activate

Step 3: Install the required packages

python3.12 -m pip install -r setup/native/requirements.txt

Step 4: Start the Jupyter Lab

jupyter lab

Option 3: Conda Installation

Step 1: Install Conda

Instructions can be found here.

Step 2: Create the Conda Environment

Start conda environment with

conda env create -p ./env -f setup/native/environment.yaml

Step 3: Start the Conda Environment

conda activate ./env

Step 4: Start the Jupyter Lab

jupyter lab

Using the Accelerator Zoo

In the Jupyter lab enviroment, navigate to accelerator-zoo/workspace/notebooks/ (if using the Docker container) or workspace/notebooks/ (if using a native installation like Pip or Conda) and select an accelerator. Each notebook contains instructions specific to that notebook and accelerator.

Contributors

The TeAAL specifications presented in this repository are the combined effort of many people over the course of many years, including:

• Timor Averbuch
• Alex Dicheva
• Joel S. Emer
• Christopher W. Fletcher
• Yuxin Jin
• Nandeeka Nayak
• Toluwanimi O. Odemuyiwa
• Michael Pellaeur
• Jules Peyrat
• Chenxi Wan
• Yingchen Wang
• Frederic Wu
• Xinrui Wu
• Yan Zhu

We would also like to thank the following people, whose insights allowed us to better understand, describe, and, in some cases, even validate our TeAAL specifications, including:

• Tae Jun Ham
• Kartik Hegde
• Tushar Krishna
• Francisco Muñoz-Martinez
• Eric Qin
• Daniel Sanchez
• Hanrui Wang
• Lisa Wu Wills
• Guowei Zhang
• Zhekai Zhang