Meta: Millions of AWS Graviton5 Cores to Support AI Growth

Meta continues to diversify its computing resources. The company led by Mark Zuckerberg has announced an agreement with Amazon Web Services (AWS) that involves the integration of tens of millions of cores based on Graviton architecture within its data centers. The processors developed internally by AWS have already found space at companies like Adobe, Apple, and Snowflake, as well as new players in the AI field like Anthropic.

The agreement, with a minimum duration of three years, represents a further step in expanding the computing capabilities necessary to support the development of advanced artificial intelligence systems, particularly in the field of so-called "agentic AI," meaning autonomous models capable of planning, reasoning, and executing complex tasks on a large scale.

While in recent years the industry's focus has mainly been on GPUs for training models, the increasing complexity of workloads is bringing CPUs back to the forefront. AWS Graviton processors, based on Arm architecture, are designed to handle a wide range of general-purpose operations and are also applicable in the later stages of training, such as large-scale inference.

Graviton5 chips are characterized by 192 cores, a cache up to 5 times larger than the previous generation, and a reduction in communication latency between cores of up to 33%. Based on the 3 nm manufacturing process, they offer up to 25% more performance while also improving energy efficiency. Additionally, the company highlights a significant advantage in terms of energy efficiency, with consumption up to 60% lower than other solutions available on EC2.

The agreement with AWS is part of a broader approach adopted by Meta, which aims to combine various solutions to optimize performance with different workloads. The company continues to invest in custom solutions like MTIA accelerators (a partnership with Arm) while also turning to NVIDIA and AMD for GPUs.

Infrastructure expansion is also driven by the scale of Meta's services, used daily by about 3.6 billion people. To support this demand, the company plans to reach 32 global data centers with the opening of a new facility in Oklahoma. Meanwhile, Meta is taking measures to contain operational costs, including a reduction in the workforce of about 8,000 employees, equivalent to 10% of the total.