Meta further diversifies beyond Nvidia as it unveils four custom AI chips

By Britney Nguyen

The tech giant's custom AI chips are used to train and power its ranking and recommendations systems and AI models

Meta announced new generations of its MTIA custom AI chips on Wednesday.

Meta Platforms announced four new generations of its custom artificial-intelligence chips on Wednesday, reflecting a push to diversify its compute options.

The tech giant (META) introduced its MTIA 300, 400, 450 and 500 chips, which will be used to train and run its ranking and recommendation systems, as well as its AI applications and Llama models. Meta said some of its new chips have already been deployed, while others will roll out later this year or next year.

The new Meta Training and Inference Accelerator chips follow the MTIA 100 and 200 accelerators, which were released in 2023 and earlier this year, respectively. Meta works with Broadcom (AVGO) on its MTIA chips, which are known as application-specific integrated circuits, or ASICs.

Read on: Why Broadcom's earnings report has Wall Street so upbeat on a bad day for chip stocks

Compared with merchant graphics processing units from Nvidia (NVDA) and Advanced Micro Devices (AMD), ASICs are designed for specific uses. The first MTIA ASIC was designed in 2020 for internal workloads, according to Meta.

Supporting AI at scale while keeping costs low "is one of the most demanding infrastructure challenges in the industry," Meta said in a blog post.

"While we remain committed to a diverse silicon portfolio and to leveraging the best solutions available - both internally and externally," Meta said its custom chip family "has remained and will continue to be an important part of" its approach to AI infrastructure.

Meta announced late last month that it had made a deal with AMD to deploy up to 6 gigawatts of the chip maker's Instinct GPUs, as well as its Epyc central processing units and Helios rack-scale system. The first GPU deployment, which is expected to start in the second half of this year, is custom-designed, but based on AMD's MI450 architecture.

See more: AMD's stock rockets as Meta deal serves as major validation point for investors

Meta CEO Mark Zuckerberg said at the time that the multiyear partnership with AMD is part of its efforts to diversify compute supply. The company said in its fourth-quarter earnings report in January that it expects capital expenditures to be between $115 billion and $135 billion this year as it builds out its AI data centers.

The MTIA 300 was initially designed for Meta's ranking and recommendation models, which made up most of the company's workloads before AI, it said. While that chip is currently in production for training its ranking and recommendation systems, Meta said the following MTIA 400 was optimized for generative AI and is on track for deployment in its data centers soon.

Meanwhile, the MTIA 450 and 500 chips were developed to focus on AI inference, or running AI models after training, and are both slated to be mass-deployed next year.

Meta said it doubled the amount of high-bandwidth memory from the 400 to the 450, as that component has become "the most important factor affecting GenAI inference performance."

As AI model-context windows, or the amount of data processed at one time, increase with more powerful models, the need for memory and storage has grown exponentially. That has led to severe shortages that are expected to disrupt the consumer-electronics industry and further AI development.

Don't miss: Micron's stock climbs on this fresh sign of booming memory prices

The amount of HBM in the MTIA 450 is "much higher than that of existing leading commercial products," Meta said. The MTIA 500 chip has 50% more HBM bandwidth than the 450.

-Britney Nguyen

This content was created by MarketWatch, which is operated by Dow Jones & Co. MarketWatch is published independently from Dow Jones Newswires and The Wall Street Journal.

(END) Dow Jones Newswires

03-11-26 1349ET