Why did YouTube develop its own chip 'Argos' specialized for video conversion?

With over 500 hours of video uploaded every minute to YouTube, the world's largest video sharing platform, much of Google's data center computing is spent converting videos. Therefore, YouTube has independently developed a chip 'Argos' specialized for video conversion. Protocol of IT media summarizes how Google developed Argos.

YouTube custom Argos ASIC chips are the future for Big Tech - Protocol


Google had invested tens of billions of yen in designing an application-specific integrated circuit (ASIC) specialized for machine learning called the tensor processing unit (TPU). Google's AI-related efforts are one step ahead of its competitors due to this TPU technology.

On the other hand, while developing TPU, Google realized that AI is not the only computing that can be improved. Parthasarathy Ranganathan, Vice President of Engineering at Google, noted that YouTube video conversion is the most computationally intensive application in Google's data centers. 500 hours worth of videos are uploaded to YouTube every minute, and in order to be able to play on various devices such as smartphones, TVs, and PCs, video conversion work is being done at the data center, and 15 videos are uploaded per video. It seems that it can also be converted to different formats.

However, the chips used in Google's data centers weren't very good at video conversion. Therefore, Ranganathan thought that instead of an ASIC specialized for AI, a powerful chip that could efficiently perform the most frequently performed video conversion was needed.

Ranganathan and his colleague Danner Stodorsky sent an email to Scott Silver, vice president of YouTube, requesting a budget of about 40 people and hundreds of millions of yen. After that, we had a 10-minute meeting with YouTube's CEO Susan Wojcicki, and the go-ahead was given to the development project of YouTube's first video conversion specialized chip. “Given the economics, workload, and our work, I thought that Mr. Ranganathan's idea made a lot of sense,” said Silver.

A chip dedicated to YouTube video conversion was given a codename after a monster from Greek mythology called 'Argos'. The existence of Argos was published in April 2021, and the concept of the chip is summarized in the following article.

What is YouTube's proprietary chip 'Argos' for video conversion? -GIGAZINE

Jay Goldberg, president of IT consulting firm D2D Consulting , said, ``The purpose of saving hundreds of yen in terms of margin is not worth the enormous cost of hiring and training chip designers. Developing a new product costs tens of billions of yen just to make a prototype, and hundreds of millions of yen to complete it.' In response to these opinions, Vice President Silver commented, ``Our concern is not saving money. I'm here.

In addition, Vice President Silver said, ``When we thought about machine learning training and inference, these were really big and interesting workloads that CPUs couldn't handle well. It may be possible, but if you're paying tens of billions of yen a year for it, there's room to invest in developing an ASIC that specializes in video conversion.'

Furthermore, Protocol points out that one of the reasons why Google developed a chip specialized for video conversion is not only cost reduction, but also a major change in the chip market. In the 1990s and 2000s, dozens of companies were vying to make the chips demanded by big tech companies, and the fierce competition created many options. However, at the time of writing the article, there are only a few major chip makers such as Intel, AMD, and NVIDIA, companies that make chips for data centers. As a result, big tech companies will use generic processors made by each manufacturer instead of the chips they want. Of course, such processors are also high enough in performance, but they are not purpose-specific.

Adopting a new independently developed chip instead of a general-purpose processor requires a fundamental rethinking of the machine configuration of the data center, such as the board on which the chip is mounted, the design of the data center rack, and the configuration of each cluster. Ranganathan said, ``It's possible to develop great hardware, but if the software you use with it isn't built to actually work on the new chip, there's a lot of work involved in compiling, tooling, debugging, and deployment. I will do it,” he said.

As a result, Google has developed a simultaneous Electronic Design Automation (EDA) tool called 'Taffel' to help develop TPU and Argos processors. ``The idea of using a software-centric approach to hardware design is something we pushed very hard at Argos,'' Ranganathan said.

YouTube calls the hardware configuration including Argos VCU (Video Coding Unit) and uses it to convert 500 hours of videos uploaded every minute into various formats. The new design incorporated into Argos improves the computational performance of video conversion by 20-33 times. At the time of writing, Google has deployed second-generation Argos chips on thousands of servers around the world, and is developing two next-generation chips.

'It's not enough for Google to build one generation of VCUs that can compete with chips from NVIDIA, AMD, and Intel. For custom chip development to make sense, Google is years ahead of the semiconductor giants,' Protocol said. otherwise it would make more sense to wait for the semiconductor giants to develop a similar chip, but for YouTube it's a very good chip for one purpose It makes a lot more sense to design it and leave the rest to expensive commodity chips.'

in Web Service,   Hardware, Posted by log1i_yk