According to an internal company document viewed by Reuters on Thursday, Meta Platforms, the owner of Facebook, is set to deploy a new version of its custom chip into its data centers in 2024. The chip, the second generation of Meta’s in-house silicon line, aims to support the company’s artificial intelligence (AI) initiatives, potentially reducing its reliance on Nvidia chips and controlling the escalating costs associated with AI workloads.
Meta has been aggressively enhancing its computing capacity to support generative AI products across platforms such as Facebook, Instagram, WhatsApp, and hardware devices like Ray-Ban smartglasses. This move involves significant investments in specialized chips and data center reconfigurations. The deployment of its in-house chip could lead to substantial savings in annual energy costs and chip purchasing costs, given Meta’s vast scale.
The updated chip, internally referred to as “Artemis,” is expected to work alongside off-the-shelf graphics processing units (GPUs) that Meta is purchasing. A Meta spokesperson confirmed the plan, highlighting the complementarity of internally developed accelerators and commercially available GPUs for Meta-specific workloads.
Meta CEO Mark Zuckerberg revealed plans to have approximately 350,000 flagship “H100” processors from Nvidia by the end of the year, equivalent to a total compute capacity of 600,000 H100s when combined with other suppliers. This deployment represents a positive turn for Meta’s in-house AI silicon project, following the decision to halt the chip’s first iteration in 2022.
While the new chip, like its predecessor, is designed for inference, Meta is also reportedly working on a more ambitious chip capable of both training and inference, similar to GPUs. Despite early setbacks, an inference chip could offer greater efficiency for Meta’s recommendation models compared to the power-intensive Nvidia processors, potentially leading to significant cost savings.
