- Meta is reportedly readying its first in-house AI training chip for deployment
- The dedicated AI accelerator, made with TSMC, completed tape-out
- Meta’s shift to custom silicon aims to reduce its dependence on Nvidia hardware
Like many of Nvidia’s highest spending customers, Meta is looking to slash its reliance on the GPU maker’s expensive AI hardware by making its own silicon.
In 2024, the social media giant began advertising for engineers to help build its own state-of-the-art machine learning accelerators, and now, according to an exclusive report from ReutersMeta is at the testing stage for its first in-house chip designed for training AI systems.
Sources told Reuters that following its first tape-out of the chip, Meta has started a limited deployment, and if testing goes well, it plans to scale production for wider use.
RISC-V business
According to Reuters, “Meta’s new training chip is a dedicated accelerator, meaning it is designed to handle only AI-specific tasks. This can make it more power-efficient than the integrated graphics processing units (GPUs) generally used for AI workloads.”
Taiwan-based chipmaker TSMC produced the silicon for Meta as part of the Facebook owner’s Meta Training and Inference Accelerator (MTIA) program, something which Reuters points out has had “a wobbly start for years and at one point scrapped a chip at a similar phase of development.”
In 2023, Meta unveiled its first-generation in-house AI inference accelerator designed to power the ranking and recommendation systems for Facebook and Instagram, and then in April 2024 it debuted a new version that doubled the compute and memory bandwidth.
At the 2024 Hot Chips symposium, Meta revealed that its inference chip was built on TSMC’s 5nm processwith the processing elements on RISC-V cores.
Like a growing number of tech firms, Facebook has thrown its weight behind RISC-V in order to recognize its AI ambitions, and although the Reuters report doesn’t provide any details on the technical aspects of Meta’s new AI training chip, it seems a fair bet that it too will be based on the open source RISC-V architecture.
The Reuters article does note that Meta executives say they want to start using their own chips for training by next year.