Your submission is now in Draft mode.

Once it's ready, please submit your draft for review by our team of Community Moderators. Thank you!

Submit Essay

Once you submit your essay, you can no longer edit it.

Pending

This content now needs to be approved by community moderators.

Submitted

This essay was submitted and is waiting for review.

How much computation did OpenAI Five (Finals 2019 version) use for training?

Question

This is related to the AlphaFold, AlphaStar, GPT-2 and OpenAI Five (TI 2018 version) computation guesstimate questions.


We continue our series of questions to track the compute used by major projects (the previously most extensive tracking, OpenAI's AI and Compute paper, only tracks projects up to the end of 2017).

This question asks about computation in PFLOP/s-days used by the version of OpenAI Five that defeated world champions OG in April 2019 (match announcement here).

The estimate should not include computation used for hyperparameter tuning and architecture search.

Resolution by paper or other reliable announcement.

The method of calculations should be as similar as possible to that used in the "AI and compute" article.


Resources

As a pointer, DanielFilan and James' Guesstimate model of Starcraft timelines also does some estimates of the dollar cost of training OpenAI Five.

Here's a Guesstimate model for quickly thinking about the relation between compute and cost, and how it changes over time, based on Ryan Carey's estimates. I especially recommend checking out the calculators (by pressing the little calculator symbol in the menu bar), which allow you to enter a target compute/cost and find out how long until it's reached.

Make a Prediction

Prediction

Note: this question resolved before its original close time. All of your predictions came after the resolution, so you did not gain (or lose) any points for it.

Note: this question resolved before its original close time. You earned points up until the question resolution, but not afterwards.

Current points depend on your prediction, the community's prediction, and the result. Your total earned points are averaged over the lifetime of the question, so predict early to get as many points as possible! See the FAQ.