- Meta is reportedly ready for its first internal AI training chip for implementation
- 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 consumption customers, Meta looks at cutting its dependence on the GPU manufacturer’s expensive AI hardware by producing its own silicon.
By 2024, the social media giant began advertising for engineers to help build their own advanced machine learning accelerators, and now, according to an exclusive report from PakinomistMeta is at the test stage for its first internal chip designed for training AI systems.
Told sources Pakinomist The fact that Meta after its first tape out of the chip has started a limited implementation and if the test goes well, it plans to scale production for wider use.
RISC-V business
According to Pakinomist, “Meta’s new training chip is a dedicated accelerator, which means it is designed to handle only AI-specific tasks. This can make it more efficient than the integrated graphics treatment units (GPUs) generally used for AI workload.”
Taiwan-based Chipmaker TSMC produced silicon for Meta as part of the Facebook owner’s meta-training and inference accelerator (MTIA) program, something like Pakinomist Points have had “a Wobbly start for years, and at one point a chip scrapped in a similar phase of development.”
By 2023, Meta revealed its first generation’s internal AI-inferencer accelerator designed to operate the ranking and recommendation systems for Facebook and Instagram, and then in April 2024 it debuted with a new version that doubled the calculation and memory width.
At Symposium in 2024, Meta revealed that its inferens chip was built on the TSMC’s 5NM process with the treatment elements of RISC-V-Kernes.
As a growing number of tech companies, Facebook has thrown its weight behind RISC-V to acknowledge its AI ambitions, and although Pakinomist Report provides no details of the technical aspects of Meta’s new AI training chip, it seems to be a fair venture that it will also be based on the Open Source RISC-V architecture.
The Pakinomist Article notes that meta -managers say they will start using their own chips for training by next year.
