Shulou(Shulou.com)11/24 Report--

[introduction] Meta faces a thorny problem: despite the huge amount of money spent on AI research, the progress of product transformation is slow, and it was not until ChatGPT became popular that it began to pay attention to it. The more expensive chip deployment that Meta previously needed for generative AI, such as GPU, was not enough, but depended on CPU. When it turned out to be unworkable, Meta turned to the multibillion-dollar Nvidia GPU, but had been dumped by giants such as Microsoft and Google. Currently, Meta plans to develop a new chip that can train AI models and perform reasoning like GPU, which is scheduled to be completed around 2025. In addition, the person in charge also said that Meta obviously needs to improve the tools and processes developed by AI.

Meta's internal emails show that in the late summer of 2022, Meta CEO Mark Zuckerberg (Mark Zuckerberg) summoned his main aides to analyze and discuss Meta's computing power for five hours, focusing on Meta's ability to carry out advanced artificial intelligence (AI) work.

According to internal emails, company announcements and people familiar with the matter, Meta faces a thorny problem: despite massive investment in AI research, Meta has made slow progress on how to introduce AI-friendly software and hardware systems into its main business. At a time when Meta is increasingly relying on AI to support further growth, this has affected the pace at which the company promotes comprehensive innovation.

"We are obviously lagging behind in terms of tools, workflows and processes in the development of AI, and we need to make a lot of investment in this area," according to the email from Santosh Santosh Janardhan, Meta's new head of infrastructure. The email was posted on Meta's intranet in September and was recently exposed for the first time.

The email also said that supporting AI requires Meta to "fundamentally change the design of hardware infrastructure, software systems, and the way in which it provides a stable platform."

For more than a year, Meta has been working on a huge project to complement its AI infrastructure. Although Meta publicly admits to lagging behind in AI's hardware development, details of this, including computing pressure, management change and an abandoned AI chip project, have never been reported before.

In response to the email and related restructuring measures, Meta spokesman Jon Cavell (Jon Carvill) said that "with its deep expertise in AI research and engineering development, Meta has been proven in building and deploying the most advanced large-scale infrastructure".

"as we provide new AI experiences for applications and consumer products, we are confident that we can continue to expand our infrastructure capacity to meet short-and long-term needs," he said.

But he declined to respond to news that Meta had abandoned its AI chip project.

The restructuring led to a quarterly increase in capital expenditure of about $4 billion, almost double that of 2021, and led to the suspension or cancellation of plans to build data centres in four locations, according to Meta.

Meta is also under financial pressure. Since November, Meta has launched large-scale layoffs that have not been seen since the collapse of the internet in the millennium.

On the other hand, Microsoft-backed OpenAI released ChatGPT on November 30 last year, an AI chat robot that quickly became the fastest-growing consumer app in history and sparked an AI arms race among tech giants. Large technology companies have launched their own generative AI products. In addition to recognizing patterns in the data, this AI can also generate text and visual content in a human-like way based on input information.

Several sources say that generative AI consumes a lot of computing power, which makes Meta more urgent to expand its computing infrastructure.

1. According to sources who have not invested enough in expensive projects, one of the crux of the problem is that Meta started introducing GPU chips into AI work very late. GPU is very suitable for AI computing, which can perform a large number of tasks in parallel and greatly reduce the time-consuming of processing large amounts of data. Of course, the price is also more expensive, with Nvidia accounting for 80% of the market share.

Therefore, Meta largely relies on CPU to undertake the task of AI computing. CPU is the main chip in the computer industry, which has spread all over the world in data centers in the past few decades, but it is not suitable for dealing with AI computing tasks.

According to two sources, Meta also uses self-designed custom chips for AI reasoning. However, by 2021, the adoption of CPU and customized chips has proved to be slower and less efficient than GPU in the field of AI. In addition, GPU is also more flexible when running different types of AI models than the chips used by Meta.

Meta declined to comment on the performance of its AI chip.

Sources said that as Zuckerberg pushed Meta to meta-universe, computing pressure affected Meta's ability to deploy AI to deal with competitive threats, such as the rise of social media rival TikTok and Apple-led changes to advertising privacy policy.

These setbacks have also attracted the attention of Peter Thiel, a former Meta board member. He resigned from Meta's board in early 2022 without giving any explanation.

At a board meeting before he resigned, Mr Thiel told Meta executives that they were too complacent about Meta's core social media business and too obsessed with Meta, according to people familiar with the matter.

2. Opting for GPU instead, but lagging behind a source, Meta executives began buying the multibillion-dollar Nvidia GPU in 2022 after canceling plans for a large-scale deployment of custom reasoning chips. By this time, Meta had lagged behind competitors such as Google. Google started deploying a customized version of GPU, or TPU, as early as 2015.

In the spring of 2022, Meta executives also began restructuring Meta's AI division, appointing two new engineering leaders, including Janalhan, the author of the September email. According to LinkedIn and people familiar with the matter, more than a dozen managers left Meta during a period of turmoil that lasted for months. The management team of the MetaAI infrastructure has been almost completely replaced.

Next, Meta began to redesign the data center infrastructure to accommodate the upcoming deployment of GPU chips. Compared with CPU,GPU, it consumes more power and generates more heat, and it needs to connect a large number of chips through a specially designed network to form a cluster.

According to Janalhan's email and sources, these facilities require 24 to 32 times the network capacity, as well as a new water cooling system to manage the cooling of the chip cluster, so the facilities need to be "completely redesigned."

As the work progressed, Meta made an internal plan to develop a new autonomous chip. The chip, which trains AI models and performs reasoning like GPU, is currently scheduled to be completed around 2025.

Carville, a spokesman for Meta, said that some data center construction projects are currently suspended and will transition to new designs, which will be restarted later this year. He declined to comment on Meta's internal chip project.

3. The slow progress of product landing in the process of expanding the computing power of GPU, Meta has almost no new product technology to show. By contrast, companies such as Microsoft and Google are promoting the open commercial use of generative AI products (Bing chat, Bard, etc.).

In February, Susan Li, Meta's chief financial officer, admitted that he was not devoting much computing power to generative AI. "basically all of our AI capabilities are given to advertising, information streaming and short video Reels," she says.

According to sources, it was not until the launch of ChatGPT in November last year that Meta began to pay attention to generative AI products. They say that although FAIR, Facebook's AI lab, has been releasing prototypes of the technology since the end of 2021, it has not turned research into a product.

With the increase in investor interest, that is changing. In February, Zuckerberg announced the creation of a top-level generative AI team that would "significantly boost" the company's work in this area.

Andrew Bosworth, chief technology officer of Meta, also said this month that generative AI is the area where he and Zuckerberg spend the most time so far, and a related product is expected to be released this year.

Two people familiar with the new team said that the team's work is at an early stage, focusing on building the basic model as the core, which can be adjusted to different product needs in the future.

Meta spokesman Cavell said that for more than a year, many Meta teams have been developing generative AI products. He confirmed that work in this area had accelerated in the months since the arrival of ChatGPT.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.