Meta has unveiled a new suite of four custom-designed chips, the latest iteration of its Meta Training and Inference Accelerator (MTIA) family, specifically engineered to bolster its artificial intelligence infrastructure. This strategic move by the social media giant signals a continued commitment to in-house silicon development, aiming to optimize performance, cost-efficiency, and supply chain diversification in its burgeoning data center operations.
The newly announced chips are part of Meta’s ambitious expansion plans, which include massive data center developments like the 5-gigawatt Hyperion facility under construction in Louisiana. These custom chips, manufactured by Taiwan Semiconductor (TSMC), are a critical component of Meta’s strategy to gain more leverage in its capital expenditure and to mitigate the volatility of external chip markets.
Yee Jiun Song, Meta’s Vice President of Engineering, emphasized the strategic advantages of this custom silicon approach. “By designing custom chips, which are then manufactured by Taiwan Semiconductor, the social media giant can squeeze more price per performance across its data center fleet rather than relying on only vendors,” Song told CNBC. He added, “This also provides us with more diversity in terms of silicon supply, and insulates us from price changes to some extent. This is a little bit more leverage.”
The MTIA family, first publicly revealed in 2023, has seen rapid iteration. The first of the new chips, MTIA 300, has already been deployed and is optimized for training smaller AI models that power Meta’s core ranking and recommendation engines across its applications like Facebook and Instagram. These models are crucial for delivering personalized content and targeted advertising.
Looking ahead, three more advanced chips—MTIA 400, MTIA 450, and MTIA 500—are slated for release. The MTIA 400 has successfully completed its testing phase and is on the cusp of deployment, designed to accelerate AI inference tasks, particularly those involving generative AI such as creating images and videos from textual prompts. The MTIA 450 and MTIA 500 are slated for operational readiness in 2027 and will also focus on advanced inference workloads, though Meta has clarified they will not be used for training exceptionally large language models.
The rapid cadence of chip development is a deliberate response to Meta’s aggressive scaling of its AI infrastructure. “It’s unusual for any silicon company or team to be releasing a new chip every six months. It’s a very quick cadence,” Song explained. “And the big reason for this is that we find ourselves building out capacity so quickly at the moment, and spending so much on CapEx, that at any given time we want to have the state-of-the-art chip to deploy.” Meta anticipates these chips will maintain a “standard five-plus years of useful lifetime.”
Meta’s significant investment in AI infrastructure extends beyond its custom silicon. The company is actively expanding its data center footprint with major facilities in Ohio and Indiana, in addition to the Louisiana project. Reports also indicate Meta is exploring leasing space at the Stargate data center site in Texas, a move that could be influenced by the recent withdrawal of OpenAI and Oracle from expanding their presence there.
This trend of in-house chip development is not unique to Meta. Major technology firms like Google, with its Tensor Processing Units (TPUs) since 2015, and Amazon, with its Inferent chips announced in 2018, have been pioneering custom silicon. These “hyperscalers” are increasingly turning to Application-Specific Integrated Circuits (ASICs) as a more cost-effective and specialized alternative to the ubiquitous, but often supply-constrained and expensive, Graphics Processing Units (GPUs) from industry giants like Nvidia and AMD. While Google and Amazon offer their custom chips to customers via their cloud platforms, Meta’s MTIA chips are exclusively for internal use.
The development of these advanced chips, particularly those intended for generative AI inference, necessitates significant amounts of high-bandwidth memory (HBM). This requirement places Meta’s ambitious roadmap squarely in the path of current industry-wide memory chip shortages. “We’re absolutely worried about HBM supply,” Song admitted. “But we think that we have secured our supply for what we’re planning to build out.” The memory market is notoriously cyclical, with chipmakers typically relying on short-term contracts with suppliers like Samsung, SK Hynix, and Micron. While Song declined to comment on specific long-term contracts, he affirmed Meta’s “diversified” approach to its supply chain and silicon strategy.
In parallel with its custom silicon efforts, Meta has recently secured substantial partnerships for external GPUs, including multi-year deals for millions of Nvidia GPUs and up to 6 gigawatts of AMD GPUs. “The workloads are changing so quickly that we want to make sure that we have options,” Song stated, underscoring the need for flexibility in their hardware strategy.
The manufacturing of Meta’s in-house silicon is handled by TSMC, a critical player in the global semiconductor supply chain, with major operations in Taiwan and a burgeoning fabrication campus in Arizona. Meta has not disclosed whether its new MTIA chips will be produced at the Arizona facility. A significant portion of Meta’s engineering talent dedicated to silicon development, numbering in the hundreds, is based in the United States, aligning with the company’s substantial domestic data center presence, with 26 of its 30 operational and planned data centers located within the U.S.
Original article, Author: Tobias. If you wish to reprint this article, please indicate the source:http://aicnbc.com/19734.html
