Meta Reveals New Details for In-House Chip Designed to Make AI Apps and Data Centers More Efficient

When it comes to massive tech companies, in-house chip investments have become rather common. From the likes of Google and Apple to now Meta, making internal chips to power global operations is no longer novel. But it can still be impactful.  

Meta recently shared new details about its projects to build a custom chip “family” to power its data centers and artificial intelligence projects that could arrive as soon as 2025.

The company released a series of blog posts describing its efforts to create a first-generation chip through its Meta Training and Inference Accelerator (MTIA) program in 2020. Previous reports claimed that the company doesn’t intend to widely roll out its first in-house chip. Rather, it is using the process as a learning experience to guide its development of new chips in the future.  

Inferring Better Results

Meta currently operates 21 data centers around the world to power Facebook, Instagram, Messenger, and its lesser-known services. Covering 40 million square feet, these campuses account for $16 billion of investment in both research and hardware.  

With data centers taking up so much of the company’s operations, it makes sense that Meta wants to make them more efficient. Few companies operate at a level where investing in custom chip production is beneficial. But Meta is certainly one of them.  

The first MTIA chip was designed to focus on improving the artificial intelligence (AI) process known as inference. Used across almost all social media platforms today, inference algorithms decide what content a user sees next. At scale, these algorithms play a huge role in the popularity of apps like Facebook and Instagram, which constantly update a user’s feed to match their interests.  

The new chip uses just 25 watts of power—significantly less than those offered by Nvidia and other chipmakers. It also operates on an open-source architecture called RISC-V, which has garnered significant interest within the chip world over the past decade. The MTIA chip, combined with GPUs, allows Meta to deliver “better performance, decreased latency, and greater efficiency for each workload.”  

According to Meta’s recent blog posts, the company originally tried using GPUs to handle its inference tasks. But it quickly found that they were not well-suited to the work and began looking for a new solution.  

Meta software engineer Joel Coburn said, “Their [GPU] efficiency is low for real models, despite significant software optimizations. This makes them challenging and expensive to deploy in practice. This is why we need MTIA.”  

Patience is Key

Inference isn’t the only area Meta is targeting with its in-house chip efforts, though. The company is also in the midst of a massive project to improve its overall AI infrastructure as the technology becomes more ubiquitous across the social media landscape.  

Unfortunately, the first MTIA chip lacked the necessary hardware and software support needed by teams building AI features beyond inference. So, while Meta originally planned a large-scale rollout of the first chip in its data centers, it is now taking a more patient approach.  

Now, Meta is working on a second, more ambitious chip. The company’s blog posts noted that it will be able to handle both inference tasks and low-to-medium complex AI tasks more efficiently than generic chips on the market today. The second-generation hardware isn’t set to begin rolling out until 2025.  

This chip will play an important role in Meta’s plans for new data centers moving forward. The social media giant aims to use updated AI-oriented networking and cooling systems to pave the way for such advancements. Its proposed design will reduce costs by 31% and allow Meta to build new data centers in half the time that it currently takes.  

Meanwhile, Meta is also developing (and using) an AI-powered system on the software side of its operations. Like Google, Amazon, and Microsoft, its platform allows software engineers to develop code much faster than doing it manually.  

Moving forward, there’s no doubt Meta will lean heavily on its AI-focused chip as it plans to “introduce AI agents to billions of people in new ways that will be useful and meaningful.”  

With demand for AI applications soaring, the MTIA chip could be a huge step forward for Meta. But whether Meta can get its chip into real-world data centers before competitors like Nvidia make something better remains to be seen.