In a significant ramp-up of its hardware ambitions, Meta is preparing to launch four new generations of its custom AI chips within the next two years. This accelerated timeline, far quicker than the industry's typical pace, is designed to power the company's expanding needs in ranking, recommendations, and generative AI.

The effort centers on the Meta Training and Inference Accelerator (MTIA), first introduced in 2023. Hundreds of thousands of these chips are already deployed, handling inference for content and ads across Meta's apps. By designing its own silicon for specific tasks, Meta claims it achieves better performance and lower costs than using off-the-shelf alternatives.

The new chips, designated MTIA 300 through 500, promise major gains in computing power and efficiency. The MTIA 300 is already in production for training ranking models. The subsequent generations will be versatile but are primarily earmarked to manage the expected surge in generative AI inference through 2027. A key design feature allows these new chips to slot directly into existing data center racks, speeding deployment.

Meta's strategy hinges on three principles: speed, a focus on inference, and standardization. The company has structured its development to release a new chip variant roughly every six months, a rapid iteration meant to keep pace with AI advances. Unlike chips built first for massive AI training, later MTIA models are optimized initially for the efficiency demands of running generative AI, then adapted for other tasks.

Furthermore, the chips are built on common industry frameworks like PyTorch and Open Compute Project standards, ensuring they work within Meta's broader infrastructure, which also incorporates silicon from other vendors. The company argues this portfolio approach—mixing custom and commercial chips—is necessary to meet diverse technical demands and maintain its aggressive innovation cycle.

Source: Facebook