This is how YouTube’s Argos chips transcode the videos we upload.

About seven years ago, Partha Ranganathan realized that Moore’s Law was dead. This created a big problem for Google’s VP of engineering: he concluded that chip performance would double every 18 months without increasing costs.

So much so that helped organize the payment of billions of dollars that Google spend each year on IT infrastructure around this idea.

But Ranganathan got the chip was doubling every four years, and it seemed that in the near future this gap would become even larger. So he and Google decided to do something about it.

The company has already spent hundreds of millions of dollars developing its own artificial intelligence chips, now known as Tensor processors, that are built into Google Pixels.

But when Google developed the TPU, the company realized that AI wasn’t the only type of computing that could be improved. Looking back, they have seen that the most resource-intensive applications in their data centers are related to video.

YouTube is by far the largest consumer of video-related computing in Google, but the type of chips it used to download, convert, and play billions of videos on its platform wasn’t particularly good.

To solve this problem and improve the efficiency of your servers, the team responsible for Google needed leadership to green-light the project in 2016. And they asked about 40 employees and an undisclosed sum of a million dollars to make it happen.

YouTube calls this chip the Argos Video Encoding Unit (VCU), and its main job is to convert the 500 hours of videos that are uploaded to the site every minute. in various screen formats and compression required to play videos on any device.

Sometimes this means that there are up to 15 variations of each video. Although the purpose of the chip was simple, and Ranganathan and the engineering team had a clear idea of ​​what they wanted to achieve, inventing a piece of silicon that would fit all needs was no easy task.

The scale needed to launch YouTube was a huge challenge which got the team thinking about design, from the chip itself to the layout of the boards they were attached to, the layout of the racks in the data centers, and the configuration of each cluster.

Learn more about how to integrate Argos chips into enterprise data centers and make them work at YouTube scale.This required close collaboration between software and hardware engineers.

As a result, Argos is hardware defined by software. This meant that the engineers working on the chip could use so-called high-level synthesis techniques to significantly speed up design iteration.

In this way, Google stopped buying chips from third parties and spending huge amounts of money to improve the efficiency and cost-effectiveness of its data centers and improving the internal workings of YouTube, a service used by millions of people every day.