Announcing a new custom chip, Meta bets heavily on AI and spends record amounts on updating infrastructure

News on May 19th: Over the past year, social media giant Meta has been committed to cutting costs and laying off employees, but it has spent record amounts updating its infrastructure in order to compete in the AI ​​​​race. maintain a leading position.

On Thursday local time, Meta released a series of new technologies, including a new custom chip that can help train AI faster, and a programming tool that provides programmers with the tools to develop products. suggestion. In addition, the company is transforming its data center to enable more convenient deployment of AI technology.

Meta CEO Mark Zuckerberg said in an emailed statement: "This work reflects our long-term efforts and will enable us to be better in every field We are developing a custom accelerator chip to speed up Facebook and Instagram’s recommendation algorithms. Additionally, the company launched a new data center designed specifically for AI hardware. Meta announced that it has completed the construction of the second phase of its AI supercomputer, which is used to train large language models such as ChatGPT.

In 2018, Meta’s capital expenditures reached US$31.4 billion, an increase of more than 4.5 times from 2017 and a record high. Zuckerberg once called this year Meta's "year of efficiency." Analysts estimate that Meta's investment level this year will be similar to that in 2022, and many of the funds will be used to improve and expand AI infrastructure.

Meta AI research director Kim Hazelwood (Kim Hazelwood) said: "While we have encountered some difficulties in improving efficiency, this does not mean that there is a direct connection between investing in AI and improving efficiency. competition.”

Some of the AI ​​updates have apparently improved Meta’s internal efficiencies, which has led to the company laying off thousands of employees in recent months.

CodeCompose is a new development tool based on generative AI that can automatically complete programming work or provide suggestions for modifying code for developers. Meta said that 5,200 internal programmers are currently using CodeCompose, and 22% of the coding suggestions made by AI are adopted.

Meta is also increasingly using AI to solve the biggest business problems. Since Apple's privacy policy adjustments have made it more difficult for many advertisers to accurately target users with their digital ads, Meta plans to use AI to guess users' interests. In order to compete with short-video platform TikTok, Facebook and Instagram have also begun to display content from accounts that users do not follow, which requires an algorithm to guess what users may be interested in.

Angelo Zino, an analyst at market research firm CFRA Research, said in an interview that investors will look for direct evidence of these improvements to justify the large expenditures.

Zino said that it is obvious that these things will take a certain amount of time and cannot be done in one step. This is aimed at Meta’s overall increase in capital expenditures. "There's definitely going to be a lot of review to make sure that we can accelerate some of the returns in terms of revenue."

Specific types of computational processes are used to make inferences, and when we ask AI models, they give answers. Meta decided to develop a new chip MTIA (Meta Training and Inference Accelerator, meaning Meta training and inference accelerator) to help complete specific internal tasks and supplement the insufficient number of NVIDIA GPUs (even if they have been purchased in large quantities).

Meta hopes that the MTIA chip will improve its ability to predict more accurate and interesting types of original and advertising content for users, so that people will spend more time using apps and click on more ads.

The company has also launched an in-house specialized integrated circuit (ASIC) designed specifically for processing video and live content.. More than 2 billion short videos are shared on Facebook and Instagram every day. With this new processor, videos will load faster and consume less data, no matter what device they are watching on.

Meta’s VP of Hardware Engineering Alexis Bjorlin said: “We successfully optimized, balanced and fine-tuned the first chips to fit our recommendations model. We already have a full grasp of the information involved in generative AI workloads or other related needs.” Meta’s AI supercomputer Research SuperCluster. The company will use this supercomputer to train a large number of AI models.

Meta announced on Thursday that it has completed the second phase of construction and trained a large language model called LLaMA. This will be a key part of how the company builds its metaverse of virtual reality platforms.

Meta has been committed to opening up some of its advanced technologies to the outside world. While much of the hardware in its stack is restricted to internal use, some of the work is open source. Like the LLaMA model, the AI ​​model trained by the Meta supercomputer and capable of solving ten International Mathematical Olympiad competition problems has also been shared with researchers. CodeCompose is a code development tool built based on the technical results disclosed by the Meta AI research team. In addition, the new inference chip will help Meta continue to support PyTorch, the open source AI framework it created and has moved to the Linux Foundation to increase its independence.

Although Meta has been developing AI tools for years, Zuckerberg has chosen to base the company's future on a vaguer vision of virtual reality. Scott Kessler, an analyst at investment research firm Third Bridge, said that this shift has been severely scrutinized by investors, so this deep investment in AI infrastructure can help investors rebuild their confidence in Zuckerberg. confidence in the overall strategy.

As the industry races to integrate AI into their businesses, Kessler said, “They don’t want to lose out at the starting line,” he said. “More people believe in this technology now than they did six to nine months ago.” Say it." (Chenchen)

The above is the detailed content of Announcing a new custom chip, Meta bets heavily on AI and spends record amounts on updating infrastructure. For more information, please follow other related articles on the PHP Chinese website!