Post 30: How much did it cost to train AlphaFold2? 💸

Recently, the preprint of OpenFold was published, which is an “optimized clone” of AlphaFold2 that emerged as an open-access project in case AlphaFold2 was not released in this way. Among many interesting things, it is mentioned that the training of OpenFold cost 50,000 hours of GPU computing (Tesla A100), or in other words, between 150,000 to 200,000 dollars depending on the cloud service used (Google or Azure).

Although DeepMind has an agreement with Google and both algorithms are not strictly equal (among other things), we can approximate that ONLY the training of AlphaFold2 could have cost around 200,000 dollars, or in other words, 2,285,949 riko-pollo cubes. So many cubes for an algorithm that changed science and drove and will drive incredible biotechnological advancements.

What else could you buy with 200,000 dollars? 🤔

Refs:

1. OpenFold: Retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization