Facebook parent company, Meta Platforms has revealed details about the next generation of its in-house artificial intelligence (AI) accelerator chip, aiming to decrease its reliance on Nvidia’s AI chips and cut energy costs.
Reuters previously reported Meta’s plan to deploy a new version of a custom data centre chip, internally named “Artemis,” to tackle the increasing demand for computing power in Facebook, Instagram, and WhatsApp’s AI products.
The new Meta Training and Inference Accelerator (MTIA) chip, produced by Taiwan Semiconductor Manufacturing Co on its “5nm” process, promises three times the performance of its predecessor.
“This chip’s architecture is fundamentally focused on providing the right balance of compute, memory bandwidth, and memory capacity for serving ranking and recommendation models,” the company wrote in a blog post.
Meta’s CEO, Mark Zuckerberg, highlighted the company’s strategy to diversify its chip suppliers, intending to acquire roughly 350,000 flagship H100 chips from Nvidia and equivalent chips from other suppliers, totalling around 600,000 H100 chips this year.
The deployment of the new chip in Meta’s data centres is underway, with the company already utilizing it to serve AI applications. Additionally, Meta is exploring ways to expand the scope of MTIA, including support for generative AI workloads.
Beyond hardware development, Meta has made significant investments in software to optimize its infrastructure efficiently.
