meta built custom computer chips to assist with artificial intelligence and video processing tasks, and spoke publicly about them for the first time.
The social networking giant first unveiled its internal silicon chip project to reporters earlier this week ahead of Thursday’s virtual event to discuss investments in its AI technology infrastructure.
Investors eye Meta’s investment in AI and related data center hardware as it embarks on a ‘year of efficiency’ that includes at least 21,000 job cuts and major cost cuts .
It’s expensive for companies to design and build their own computer chips, but Meta believes the increased performance justifies the investment, Alexis Bjorlin, vice president of infrastructure, told CNBC. said. The company is also overhauling its data center design to focus on energy-efficient technologies such as liquid cooling to reduce excess heat.
One of the newer computer chips, the Meta Scalable Video Processor (MSVP), is used to process and send video to users while reducing energy requirements. Bjorlin said there was “nothing on the market” that could handle the task of processing and delivering four billion videos a day as efficiently as Meta wanted.
The other processor is the first in the company’s family of chips, the Meta Training and Inference Accelerator (MTIA), aimed at helping with various AI-specific tasks. The new MTIA chip specifically handles “inference” when an already trained AI model makes a prediction or takes an action.
Bjorlin said the new AI inference chip will help power some of Meta’s recommendation algorithms used to surface content and ads in people’s news feeds. He declined to say which company makes the chip, but said in his blog post that the processor is “manufactured on the TSMC 7nm process,” with chip giant Taiwan Semiconductor Manufacturing Co. suggests that it produces
He said Meta has a “multi-generational roadmap” for its AI chip family, including processors used for AI model training tasks, but didn’t provide any details other than the new inference chip. Reuters previously reported that Meta canceled one of its AI inference chip projects and started another that was supposed to roll out around 2025, but Bjorlin declined to comment on the report.
Meta isn’t in the business of selling cloud computing services like companies like Google parent Alphabet and Microsoft, so it doesn’t need to publicly talk about its internal data center chip projects, he said.
“If you look at what we’re sharing, the first two chips we’ve developed, you definitely get a little idea of what we’re doing under the hood,” says Bjorlin. said Mr. “We didn’t need to advertise this, we didn’t need to advertise this, but you know, the world is interested.”
Meta’s vice president of engineering, Aparna Ramani, said the company’s new hardware was developed to work effectively with its homegrown PyTorch software, which allows third-party developers to create AI apps. It has become one of the most popular tools used to
The new hardware will eventually be used to power metaverse-related tasks such as virtual and augmented reality, as well as the burgeoning field of generative AI, which generally refers to AI software that can create compelling text, images, and videos. It is scheduled to be
Ramani also said Meta has developed a generative AI-powered coding assistant to make it easier for its developers to create and operate software. The new assistant is similar to Microsoft’s GitHub Copilot tool, which it released in 2021 in partnership with AI startup OpenAI.
Meta also said it has completed building the second or final stage of a supercomputer called the Research Supercluster (RSC), which the company detailed last year. Meta used his 16,000 Nvidia A100 GPU-powered supercomputer for things like training his LLaMA language model for the company.
Ramani said Meta continues to act on the belief that it should contribute to open source technology and AI research to advance the tech sector. The company claims that the largest of his LLaMA language models, his LLaMA 65B, contains 65 billion parameters and is trained with his 1.4 trillion tokens pointing to the data used for AI training. made it clear that
Companies like OpenAI and Google don’t publish similar metrics for competing large-scale language models, but CNBC reported this week that Google’s PaLM 2 model has been trained on 3.6 trillion tokens and 340 billion. Reported to contain parameters.
Unlike other technology companies, Meta released the LLaMA language model to researchers so they can learn from technology. However, the LlaMA language model has since been leaked to the wider public, leading to many developers building apps incorporating the technology.
Ramani said Meta is “still looking into all things open source collaboration. Certainly we would like to reiterate that our philosophy remains open science and cross-collaboration.”
clock: AI will greatly influence the sentiment of big tech companies
