Facebook owner, the goal is testing its first internal chip for training artificial intelligence systems, an important milestone as it seeks to reduce dependence on external suppliers such as Nvidia, two sources told Reuters.
The largest social media company in the world has started a small chip deployment and plans to increase production for large scale use if the test is successful, the sources said.
The effort to develop chips internally is part of a long -term goal plan to reduce its huge infrastructure costs as the company makes high AI tools bets.
The goal, which also owns Instagram and WhatsApp, foresees total expenses of US $ 114 billion to US $ 119 billion in 2025, including up to US $ 65 billion (about R $ 380 billion) in Capital Expenses, mainly driven by AI infrastructure spending.
One of the sources said the new goal training chip is a dedicated accelerator, which means it was designed to deal with AI -specific tasks only. This can make it more energy efficient than integrated graphic processing units (GPUS), usually used for AI workloads.
The goal is working with TSMC chip manufacturer based in Taiwan to produce the chip, this source said.
The implementation of the test began after the goal concluded its first tape-out of the chip-a significant marker of successful silicon development work, which involves sending an initial project by a chip factory, the other source said.
A typical tape-out costs tens of millions of dollars and takes three to six months to complete, without guarantee that the test will be successful. A failure would require the goal to diagnose the problem and repeat the tape-out step.
The goal and TSMC declined to comment.
The chip is the latest in the company’s Meta Training and Inference (MTIA) series of the company. The program had an unstable start and at one time discarded a chip at a similar phase of development.
Last year, however, the goal began to use a MTIA chip to perform inference – the process involved in the execution of an AI system while users interact with it – in recommendation systems that determine which content appears in Facebook and Instagram news feeds.
Goal executives said they want to start using their own chips by 2026 for training, ie for the intensive computing process to feed the AI system with large amounts of data to “teach it” how to perform their functions.
Like the inference chip, the goal of the training chip is to start with recommendation systems and then move on to general AI products, such as Chatbot Meta AI, executives said.
“We are working on how we would do training for recommendation systems and eventually how we would think about Training and Inference for Generative AI,” Morgan Stanley’s conference at the Technology, Media and Telecommunications conference said last week.
Cox has described the target’s chip development efforts as “a kind of walking, crawling and running situation” so far, but said executives have considered the first -generation inference chip for recommendations a “big success.”
Previously, the goal has given up a personalized internal inference chip after it fails in a small -scale test deployment, similar to what is now doing for the training chip.
The Social Media company has been on Nvidia’s largest customers since, accumulating a GPUS arsenal to train its models, including recommendations and ad systems and its Llama models series. The units also perform inference for over 3 billion people who use their applications every day.
The value of these GPUs was questioned this year, as IA researchers are increasingly expressing doubts about how much progress can be done by continuing to “expand” large language models by increasingly adding data and computer power.
These doubts were reinforced by the launch in late January of new low -cost models of Chinese startup Deepseek, which optimize computational efficiency by depending more on inference than most traditional models.
See also: Artificial Intelligence can generate “global annihilation,” experts say
Brazil is among the countries that most use artificial intelligence
This content was originally published in goal begins to test first internal AI training chip on the CNN Brazil website.
Source: CNN Brasil
Charles Grill is a tech-savvy writer with over 3 years of experience in the field. He writes on a variety of technology-related topics and has a strong focus on the latest advancements in the industry. He is connected with several online news websites and is currently contributing to a technology-focused platform.
