Meta Unveils New AI Chips For Data Centers

Meta has released a new generation of custom AI chips made for its expanding data center infrastructure. The chips are part of the Meta Training and Inference Accelerator family. The technology helps the company grow its AI business.

The company showed off 4 new processors that are supposed to make AI tasks run faster. These chips will power systems that use generative AI and ranking recommendations. The project is part of Meta’s bigger plan for artificial intelligence.

Source: CNBC

MTIA 300 Chip Already Deployed In Production

The first chip in the new line, called MTIA 300, has already been used. Engineers put the processor into use a few weeks before the official announcement. Right now, it works with smaller AI models in Meta’s services.

These models help make the recommendation systems that Facebook and Instagram use. The algorithms decide which videos, posts, and ads users see. So, better hardware makes it faster to send content.

Future MTIA Chips Will Power Generative AI Tasks

There are plans for 3 more processors: MTIA 400, MTIA 450, and MTIA 500. These chips are made to do more complicated AI inference tasks. The main focus is on generative AI apps.

The processors will be able to do things like make images and videos from written prompts. Meta said that these chips will not be able to train large language models. Instead, they will speed up inference workloads in data centers.

Recommended Article: OpenAI Acquires Promptfoo To Strengthen AI Security Systems

Meta Grows Its Strategy For Making Custom Silicon

Meta executives say that making custom silicon gives them more control over the computing infrastructure. The company’s internal chips let them get the best performance out of certain tasks. This makes you less reliant on hardware vendors outside of your company.

Making your own processors also helps keep the company safe from price changes. A lot of tech companies have been hurt by a lack of semiconductor supplies lately. Because of this, custom chips give you strategic flexibility.

Data Center Infrastructure Supports AI Growth

To support its AI goals, Meta is spending a lot of money on building data centers. In Louisiana, work is currently going on on a big building called Hyperion. More projects are planned for Indiana and Ohio.

These data centers need a lot of computing power to run AI services. Custom hardware is a big part of making things run more smoothly. The chips should keep working for more than 5 years, according to the company.

Tech Industry Increasingly Builds Custom Chips

In the past few years, a number of tech companies have started making their own AI processors. In 2015, Google came out with its Tensor Processing Unit. Later, Amazon made its own custom silicon for use in the cloud.

Application-specific integrated circuits, or ASICs, are the name for these special processors. They usually do certain jobs better than graphics processors that are made for general use. A lot of businesses use them to cut down on their reliance on GPU suppliers.

Supply Chain Concerns Remain For AI Hardware

Even though Meta has made progress, problems with the supply chain could still affect chip production. There isn’t enough high-bandwidth memory right now for AI workloads. This part is very important for advanced processors.

Companies like Samsung, SK Hynix, and Micron make these memory chips. The demand for artificial intelligence hardware in the business world is still growing quickly. So, companies are still worried about getting enough supply.