eta, the parent company of Facebook, has reportedly started testing its first in-house for artificial intelligence training.
This marks an important step for Meta’s strategy to build more of its custom silicon, enabling the company to reduce its dependence on external hardware suppliers, such as Nvidia. They are planning to integrate these systems into their infrastructure by 2026.
Several years ago, Meta was actually one of the first organizations to develop its own RISC-V-based chips for training AI systems. Now, the company is finally taking matters into its own hands and creating its first in-house system – likely with support from Broadcom.
Meta’s decision may be also influenced by external threats such as potential delays from Nvidia or increasing competition in the AI world.
Although Meta has been using an earlier version of their chip – but only for running models, such as ranking and recommendation algorithms. This time, they will train them through an integrated in-house system to support Meta AI and other generative technologies.
"We're working on how would we do training for recommender systems and then eventually how do we think about training and inference for gen AI." – stated Chris Cox, Chief Product Officer at Meta.
To produce the custom chips, Meta has reportedly teamed up with TSMC, a chip manufacturer in Taiwan. According to Reuters, the testing phase commenced after the completion of Meta’s initial “tape-out” of the chip. This is a key milestone because now, the first design of the chip has already been sent to the manufacturing facility.
Chris Cox has also talked about Meta’s AI chip development efforts at the Morgan Stanley conference, explaining that it’s still "kind of a walk, crawl, run situation". But he also admitted that executives believe to be a “big success” so far.
Subscribe to our newsletter
However, this is not Meta’s first attempt at developing its own AI chip infrastructure. In 2022, they tried to build a custom inference chip, but it faced several setbacks due to a failure in its small-scale test deployment, resulting in the company reversing its strategy and spending billions of dollars on Nvidia GPUs.
The company is expected to allocate approximately $65 billion in capital expenditure this year, the majority of which will be invested in Nvidia GPUs and AI infrastructure. So, if Meta’s plan to reduce the reliance on external suppliers through custom hardware development proves successful, it could be a big win for the tech giant.
Meta’s move could also pave the way for more and more companies to invest in building their own custom hardware to cut costs and gain more control.