About

SPEED UP GREEN UP AI HACKATHON is an event focussed on making machine learning more efficient. Hackathon participants will make use of MIT's most green AI supercomputer (Satori) which has 256 Volta GPUs. Prizes will be awarded for the biggest efficiency gains during the hackathon. The event will kick off with tutorial sessions that will introduce particpants to machine learning on Satori and prepare teams for the hackathon.

Participants may bring their own projects to work on, or they can opt to work on problems that will be provided. Participants will be

Tutorials and warmup: January 28th
Hackathon: January 29th-30th
Location: Building 26-152, MIT
Prizes: Best green-up : $1000 - for the project that shows the greatest reduction in energy use using Satori
Best speed-up $1000 - for the project that shows the greatest speedup using Satori
Most innovative.. $500 - for the project with the most novel vision (independent of results)
Best Satori life hack $250 - best use of Satori's unique features or best Satori life hack
Most heroic achievement $250 - for the project that stretched the most to do something really bold (but still may be a work in progress)

Teams and Awards

We had an amazing set of teams, and great presentations from everyone on the final day. After deliberation the esteemed judges will make the following awards
Prize Team Name Team Members
Best green-up : $1000 - for the project that shows the greatest reduction in energy use using Satori IV-OCT image processing Mohammad Haft-Javaherian
Best speed-up $1000 - for the project that shows the greatest speedup using Satori Green Fakes Alex Andonian,
Camilo Fosco
Most innovative.. $500 - for the project with the most novel vision (independent of results) Multiverse Pruning Jonathan Frankle
Best Satori life hack $250 - best use of Satori's unique features or best Satori life hack - 1 CartDG Andrew Kirby
Best Satori life hack $250 - best use of Satori's unique features or best Satori life hack - 2 Oceananigans Ali Ramadhan,
Suyash Bire,
James Schloss
Most heroic achievement $250 - for the project that stretched the most to do something really bold (but still may be a work in progress) Brainy Bunch Jenelle Feather,
Kelsey Allen

All teams

Oceananigans Predicting monsoons with deep learning (hopefully)
CartDG Cartisian Discontinuous Galerkin Compressible Euler Equation Solver
The Brainy Bunch Compress deep models by learning effective receptive field sizes to speed up inference
Multiverse Pruning a new pruning method based on the lottery ticket hypothesis
Eyes and Ears Developing the computer vision models to detect the desease from ear drum images
Scaling Graphnets
Qubits vs Bits Standardize comparison of performance of quantum vs classical algos
IV-OCT image processing Intravascular optical coherence tomography multilayer image processing
Green Fakes Green up and speed up deepfake detection pipeline
Computational Linguistics Hebrew voice recognition

FAQ

  1. What is SPEED UP GREEN UP AI
    It is a machine learning hackathon that will look at optimizing energy efficiency of machine learning codes on a 256 Volta GPU cluster.
  2. Who can participate
    The hackathon is open to all members of the MIT community.
  3. When and where is the hackathon
    The hackathon will run from January 28th - January 30th and will be held in building 26 room 152. A tutorial session will introduce the hackathon on the morning of January 28th.
  4. What will I get out of it
    You will get to try out one of the largest modern GPU clusters on campus and run your machine learning training and inference codes on it. You will learn about running state of the art ML training and inference algorithms like BigGAN on hundreds of GPUs, and about the latest techniques, such as lottery ticket, for compressing giant networks to make them efficient. Cash awards will be given out after presentations for teams with the most efficiency gains and for teams with impressive scale-up and innovative approaches. Mentors will be available to help teams with advice and assistance.
  5. Will there be food.
    Yes.
  6. Who is hosting the hackathon.
    The hackathon is part of a series of events run by the MIT Research Computing Project and the MIT IBM Watson AI Lab.
  7. What can I do to get started.
    You can find out about the AI platform we will use at https://mit-satori.github.io. If you want to get a head start please feel free to contact Chris Hill ( cnh@mit.edu ) or John Cohn ( johncohn@us.ibm.com ) to get early access to the system!
  8. What else do I need to know.
    Access to https://mit-satori.github.io will require a laptop (or similar device) with a browser and a keyboard! If you want to haave access after the hackathon you will need to use your MIT (i.e. @mit.edu ) credentials. You will need to have Touchstone access configured for your device (see - https://ist.mit.edu/duo ).

Hackathon Schedule

Day Time
Jan 28 9-12 sign-up, onsite registration, on-boarding Breakfast
12-1:30 Welcome and hello Lunch
Intro to the Hackathon - C. Hill, J. Cohn
Greening your AI and computing - C. Hill
Intro to Satori - J. Cohn
Intro to TX-GAIA and other general MIT resources - L. Milechin
Mentor intros
Participants intros
1:30-3 Hands on tutorials
3-5 Mentors available, hackathon time Snacks
Jan 29 9-11 Mentors available, hackathon time Breakfast
11-12 Progress check in, green metrics check in
12-1 Mentors available, hackathon time Lunch
1-4 Mentors available, hackathon time
4-5 Progress check in, green metrics check in
Jan 30 9-1 Mentors available, hackathon time Breakfast
1-3 Team presentations Lunch
3-4 Judging and voting
4-5 Awards Snacks

Competition and judging

  1. Judging
    Hackathon groups will be judged for prizes based on several factors. Groups will self-report progress in meeting their chosen efficiency goals on the final afternoon of the hackathon. Judges will also discuss with groups their approaches and the lessons learned. The final afternoon presentations will take place between 1 and 3 on the Thursday (Jan 30th) of the hackathon and judges will discuss with groups their activities between 3-4. Prizes will be announced between 4 and 5. Judging will take into account both quantitative measures, like energy use reduction, and qualitative measures such as degree of novelty and vision, how well a group can explain how their chosen activity maps to goals of efficient and reduced resource utilization.
  2. Final Presentation Guidelines
    Groups should prepare 3-5 slides that cover
    • team members and their roles
    • what problem the group is working on and why
    • what they tried to improve efficiency
    • what ideas worked
    • what ideas did not work
    • what efficiency improvements they measured, and why they chose whatever measure used
    • what the group learned and any thoughts on future work
    Groups should aim for roughly 5 minutes of presentation and 3-5 minutes of open questions.
  3. Background
    While AI can be a force for good it also has an energy and emissions impact. Articles such as https://arxiv.org/pdf/1906.02243.pdf probably bias high to enhance awareness, nevertheless there is an underlying trend of growing energy and emissions footprint from the digital economy. The “green up” hackathon competition is geared to get people to think creatively about how to optimize the resource use of their workloads.

    The MIT facility housing the Satori computer ( https://mghpcc.org ) and the Satori computer itself ( https://mit-satori.github.io ) are both very energy efficient. In addition, MIT and Lincoln Laboratory (along with other areas universities) primarily purchase energy from hydroelectric sources for all their large computer resources in Holyoke, MA. This makes the Holyoke resources already some of the greenest large scale compute resources in academia. However, hydroelectric capacity is by nature limited, so it is also important to use these resources as efficiently as we can.

    This hackathon offers prizes for groups who can show reduced energy/resource impact relative to their starting point on a computer application. Groups can choose their own application or we will provide suggestions and codes for groups that do not have an existing problem. We will help participants get going with tools to measure resource usage beyond time to solution, including energy use, memory use, I/O and other metrics. Groups will be free to choose how to try and improve efficiency.

    Some possible ways you might try to improve efficiency include

    • first measure resource use and profile (energy, implied emissions and I/O, memory etc..)
    • algorithmic changes, e.g. calculate gradients better to converge faster
    • look for hot spots that can be streamlined/eliminated
    • compressing or optimizing inference networks and graphs
    • reduced/mixed precision but maintaining correctness
    • scheduling/checkpointing to improve system utilization
    • explore custom clock speed on GPUs (this is a thing!)
    • try a different language/library (Julia v Python v R v C etc..)

    Mentors will be available to work with groups to help them think about approaches as needed. Groups are also free to try other approaches of their choosing and mentors will be happy to help out.