Investors have been closely watching Meta’s investments into AI and related data center hardware as the company embarks on a “year of efficiency” that includes at least 21,000 layoffs and major cost cutting.
Although it’s expensive for a company to design and build its own computer chips, vice president of infrastructure Alexis Bjorlin told CNBC that Meta believes that the improved performance will justify the investment. The company has also been overhauling its data center designs to focus more on energy-efficient techniques, like liquid cooling, to reduce excess heat.
One of the new computer chips, the Meta Scalable Video Processor (MSVP), is used to process and transmit video to users while cutting down on energy requirements. Bjorlin said “there was nothing commercially available” that could handle the task of processing and delivering 4 billion videos a day as efficiently as Meta wanted.
The other processor is the first in the company’s Meta Training and Inference Accelerator (MTIA) family of chips intended to help with various AI-specific tasks. The new MTIA chip specifically handles “inference,” which is when an already-trained AI model makes a prediction or takes an action.
Bjorlin said that the new AI inference chip helps power some of Meta’s recommendation algorithms used to show content and ads in people’s news feeds. She declined to answer who is manufacturing the chip, but a blog post said that the processor is “fabricated in TSMC 7nm process,” indicating that chip-giant Taiwan Semiconductor Manufacturing is producing the technology.
She said that Meta has a “multi-generational roadmap” for its family of AI chips that include processors used for the task of training AI models, but declined to offer details beyond the new inference chip. Reuters previously reported that Meta cancelled one AI inference chip project and started another that was supposed to roll out around 2025, but Bjorlin declined to comment on that report.
Because Meta isn’t in the business of selling cloud computing services like companies including Google-parent Alphabet or Microsoft, the company didn’t feel compelled to publicly talk about its internal data center chip projects, she said.
“If you look at we’re sharing—our first two chips that we developed—it’s definitely giving a little bit of a view into what are we doing internally,” Bjorlin said. “We haven’t had to advertise this, and we don’t need to advertise this, but you know, the world is interested.”
Meta vice president of engineering Aparna Ramani said the company’s new hardware was developed to work effectively with its home-grown PyTorch software, which has become one of the most popular tools used by third-party developers to create AI apps.
The new hardware will eventually be used to power tasks related to the metaverse, such as virtual reality 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.
Ramani also said that Meta has developed a generative AI-powered coding assistant for the company’s developers to help them more easily create and operate software. The new assistant is similar to Microsoft’s GitHub Copilot tool that it released in 2021 with help from the AI startup OpenAI.
In addition, Meta said it completed the second-phase buildout, or the final buildout, of its supercomputer dubbed Research SuperCluster (RSC), which the company detailed last year. Meta used the supercomputer, which contains 16,000 Nvidia A100 GPUs, to train the company’s LLaMA language model, among other uses.
Ramani said that Meta continues to act on its belief that it should contribute to open-source technologies and AI research in order to push the field of technology. The company has disclosed that its biggest LLaMA language model, LLaMA 65B, contains 65 billion parameters and was trained on 1.4 trillion tokens, which refers to the data used for AI training.
Companies like OpenAI and Google have not publicly disclosed similar metrics for their competing large language models, although CNBC reported this week that Google’s PaLM 2 model was trained on 3.6 trillion tokens and contains 340 billion parameters.
Unlike other tech companies, Meta released its LLaMA language model to researchers so they can learn from the technology. However, the LlaMA language model was then leaked to the wider public, leading to many developers building apps incorporating the technology.
Ramani said that Meta is “still thinking through all of our open source collaborations, and certainly, I want to reiterate that our philosophy is still open science and cross collaboration.”
Watch: A.I. is a big driver of sentiment for big tech