YouTube unveils Argos, its powerful video encoding processor

Argos is not only the dog of Odysseus, a giant, an argonaut or a GPS beacon system, it is also a video encoding chip entirely developed by the YouTube teams whose second generation has just seen the day. The online video titan has encoding needs unique in the world as it has to encode 500 hours of video every minute – yes, you read that right. For this, YouTube is therefore developing VCUs, for Video Coding Unit.

And we are talking here not of a single encoding: according to the sending definition, each file sent to YouTube’s servers can be transcoded up to 15 times to be displayed on 8K televisions as well as on digital features phones. entry level showing 144 p. The task is enormous and requires a lot of resources.

One card, two chips, twenty encoding processors

If the YouTube teams have not given information on the fineness of the engraving of the chip, they have communicated some details about its structure and its use. Each Argos Gen.2 consists of a chip integrating ten video encoding cores, two teams of five cores each drawing from a pool of LPDDR (low consumption memory) type RAM. All driven by a tiny CPU housed in a corner of the chip.

YouTube has developed a PCI-E card system incorporating two Argos Gen.2 chips, passively cooled with a single radiator. The card is capable of encoding up to 20 4K60p streams in real time. Performance not only in terms of power, but also in terms of energy efficiency: the YouTube teams claiming that, compared to the previous classic solution (GPU / CPU + software), “we show energy efficiency improvements are in the range of x20 to x33“.

Lower hardware, energy and software bills

The development of a personalized chip is a complicated and complex process and represents significant financial risks. In the event of failure, it is the company which alone pays for the broken pieces, while the “off-the-shelf” chips are validated by specialized players and see their cost amortized by the sum of the customers. But in the case of YouTube, business needs are unique. And the bills are huge, both in terms of hardware, software and energy.

On the hardware side, if successful – as is the case with Apple’s M1 chip – the savings are substantial when the volume of chips is sufficient since there is no longer to pay the margin to the supplier. In terms of software too, more specifically codec licensing, YouTube needed to save money. H.264 and H.265 codecs require YouTube to pay royalties in one way or another. If YouTube still does it for compatibility reasons, the Argos Gen.2 chip also encodes in VP9, ​​copyright free and now also in AV1. Two codecs which therefore cost nothing to use, but which also save a lot of bandwidth.

Finally, it is also on the energy side that we must look. By developing a chip optimized 100% for its mission, YouTube improves the energy efficiency of its servers. Saving millions of watts at the end of the line after a year.

… and take off the AV1 codec

In addition to the total hardware and software control of its service, YouTube and Google are pushing pawns here in their confrontation against the tenors of the classic video market. Before the emergence of the video giants, Hollywood and the electronics industry ruled codecs through MPEG-like consortia. Any DVD or Blu-ray player incorporates these codecs and manufacturers (and therefore consumers) pay a license fee.

With the advent of video sharing, tech companies like Google, Facebook, Apple and others are fighting back with open source working groups to develop royalty-free codecs – an obviously self-interested move, since their license fees are very important.

But in order to push new codecs like the high performance AV1, you have to start making the movement. And the hardware integration of the codec within Argos itself will save YouTube not only licensing costs, but also storage and distribution costs (better compression).

Sources : YouTube blog via Ars technica

Full scientific article (PDF)


Please enter your comment!
Please enter your name here