Meta has long since chosen to abandon its metaverse ambitions and lean heavily on artificial intelligence. It has now made the jump to producing its own AI hardware. New internally designed AI training chips are now being tested by the social media giant.
Following the initial test, assuming the chips are successful, Meta plans to increase production of its in-house hardware and roll it out internally. The idea is to replace the company’s reliance on hardware manufacturers like Nvidia for its AI ambitions.
Meta’s bag of chips
Two sources, speaking to Reuters, said that the test is a stepping stone toward more custom-designed computing hardware from the company. It’s intended to help Meta cut costs by bringing more expenses in-house. It’s cheaper to manufacture your own AI hardware than it is to purchase from third parties.
If this sounds familiar, that’s because it’s what Apple is currently doing with its Apple Silicon. The difference here is that Meta’s chipsets are intended for internal use only.
The company’s fledgling AI hardware is currently in the “tape-out” stage, which is a limited production run that may or may not work as intended. Each run costs tens of millions (of dollars), but even this expense pales in comparison to the billions that are spent on AI compute.
Taiwan-based heavyweight TSMC is said to be developing the hardware for the Meta Training and Inference Accelerator (MTIA) programme, which recently began using a custom chip to perform inference functions for its AI systems. This new batch of in-testing hardware is expected to be used in training AI systems. If all goes (exceedingly) well, perhaps an Instagram-branded GPU isn’t too far in the future either.
