Meta Explains Compute Power Driving AI Infrastructure

Compute power is the essential capability that determines how much work a computer chip can accomplish and its operational speed. This power is measured in floating-point operations per second (FLOPS), which quantifies the number of calculations a chip can execute in one second. The scale of compute, or the number of chips that can run concurrently, is measured in gigawatts.

For instance, simple interactions with Meta AI, such as finding a vegan restaurant, involve billions of calculations performed in seconds. This process converts voice commands into text, routes them to data center servers, processes them through a large language model (LLM), and delivers the results instantly. Even basic actions like searching on Instagram necessitate extensive computation, including language understanding, query processing, data indexing, and result generation.

The processing power for these tasks is delivered by specialized chips housed within servers in data centers. Central Processing Units (CPUs) are fundamental processors enabling AI training and inference. While traditionally designed for sequential tasks, CPUs excel at managing network traffic, running application logic, and coordinating workflows across systems.

Graphics Processing Units (GPUs), initially developed for rendering graphics, are highly effective at performing thousands of calculations simultaneously. This parallel processing capability is vital for AI workloads such as training models to understand languages, recognize images, and engage in conversations, often requiring intensive, long-duration calculations.

Meta also employs custom chips, exemplified by its Meta Training and Inference Accelerator (MTIA) family of silicon. These custom processors are designed for specific workloads, optimizing efficiency for tasks like ranking, recommendations, and generative AI. MTIA chips are particularly optimized for AI inference workloads but also support training, offering flexibility and efficiency for Meta's AI infrastructure.

Meta is establishing a global network of AI-optimized data centers, each engineered to support both its AI workloads and other applications and services. This strategy involves a diversified approach to infrastructure, including sourcing silicon from various partners to match the right chips with specific computing requirements.

According to Meta Newsroom, this robust compute infrastructure is foundational to the future of AI innovation.