The new Meta chip, internally referred to as ‘Artemis’ is designed to perform the inference process,
Facebook parent, Meta, is set to upgrade its artificial intelligence (AI) capabilities with the deployment of a new, custom-designed chip within its data centres this year. This second-generation silicon, an advancement from the initial line Meta unveiled last year, is part of the company’s strategy to diminish its reliance on Nvidia’s widely used chips. This move could significantly impact Meta’s operational costs, particularly in running AI workloads crucial for the development of generative AI products across its platforms, including Facebook, Instagram, WhatsApp, and hardware devices like Ray-Ban smart glasses.
The new Meta chip, internally referred to as ‘Artemis’ is designed to perform the inference process, where models use their algorithms to make decisions and respond to user prompts. Meta is also reportedly working on a more ambitious chip capable of handling both training and inference tasks, similar to GPUs.
Despite the challenges encountered with the first generation of its Meta Training and Inference Accelerator (MTIA) program, the development of an inference chip by Meta could offer a more efficient solution for processing the company’s recommendation models than the currently used Nvidia processors.
The implementation of Meta’s proprietary chip could lead to substantial savings, potentially reducing annual energy expenses by hundreds of millions of dollars and cutting down on chip purchasing costs by billions. This development reflects the broader industry trend of tech firms investing huge in the expense of chips, infrastructure, and energy for AI applications.
The company reportedly plans to launch the updated chip in 2024, stressing that it would complement the extensive array of off-the-shelf graphics processing units (GPUs) Meta is acquiring. This approach aims to achieve the optimal balance of performance and efficiency for Meta-specific workloads.
Meta’s CEO, Mark Zuckerberg, revealed plans to acquire approximately 350,000 Nvidia “H100” processors, in demand GPUs for AI, by the end of the year. This acquisition, along with contributions from other suppliers, is expected to provide Meta with the computing power equivalent to 600,000 H100s.
The decision marks a positive direction for Meta’s in-house AI silicon project, especially after the company had previously halted its first chip iteration in 2022 in favour of purchasing Nvidia’s GPUs. Nvidia’s GPUs are currently leading in the AI training process, which requires processing vast data sets to train models.